Linking competitiveness in the banking sector with entrepreneurial activity : Evidence from dynamic panel data methods in a multi-country setting

(1)

Linking competitiveness in the banking sector with

entrepreneurial activity

Evidence from dynamic panel data methods in a multi-country setting

Abstract - This study investigates the relationship between competition in the bank-ing sector and entrepreneurial activity. Usbank-ing a self-constructed dataset, Arellano-Bond and Blundell-Bond GMM estimators are obtained. It is argued that the lack of dynamics in previous studies is incorrect. This study estimates the relationship using an autoregressive distributed lag model for which it is argued that it molds the dynamics satisfactory and that this leads to improved results. Competitiveness in the banking sector is measured by three different indicators and the results are compared. In line with earlier research, it is found that the Lerner index gives the most sound results. From this it is concluded that an increase in bank competition has a positive immediate and long run effect on the new business entry density rate. However, inference based on the results of this study should be done with caution since the dataset is relatively small and GMM results are considerably influenced by choices of the researcher.

Keywords: panel data, autoregressive distributed lag model, Blundell-Bond estimator, Arellano-Bond estimator, firm creation, bank market power, bank competitiveness, Lerner index, entrepreneurial activity, firm financing.

Thesis MSc. Econometrics

Author:

E.B.M. (Emma) Haccou 10178708 First Supervisor: Prof. Dr. J.F. Kiviet Second Supervisor: Dr. M.J.G. Bun April 25, 2017

(2)

Statement of originality

This document is written by Emma Brigitta Maria Haccou who declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

Preface

I would like to express my gratitude to my supervisor Professor Jan Kiviet for his lectures on panel data and the supportive feedback, not only on econometric funda-mentals and choices but also on small details. Furthermore, I appreciate the fact that he supported me by keeping me motivated and giving me the time and space I needed.

(3)

1 Introduction

Entrepreneurship is seen as a key element for economic growth and assumed to be vital for economic development and prosperity (Van Stel et al., 2005). Already in the beginning of the twentieth century this was emphasized by Schumpeter (1934). In his later work (Schumpeter, 1942), Schumpeter elucidated on the innovative role of entrepreneurs in particular and he saw this as the main force of economic growth. According to his research, innovations and entrepreneurs challenge established firms and hence strengthen the economy as a whole.

The contribution of entrepreneurial activity to economic growth is investigated in many papers and is examined and explained in various ways. Leibenstein (1968) stresses the important role of entrepreneurship within the development process. The positive role between entrepreneurial activity and the evolution of industries is es-tablished by Audretsch & Thurik (2003). In addition, entrepreneurs may initiate important innovations (Wennekers & Thurik, 1999), which positively affect economic growth. Furthermore, entrepreneurial activity enhances competition and productivity (Wennekers & Thurik, 1999; Nickell, 1996). Haltiwanger et al. (2013) investigate the contribution of entrepreneurial activity to job creation and highlight the important role of young firms in the labor market.

Many researchers, such as Gnyawali & Fogel (1994), underline the influence of entrepreneurial environments for business start-up and success. ˇCih´ak et al. (2012) address the issue of poor financial systems and state that this hinders economic growth and impedes entrepreneurs. Commissioned by the World Bank, the authors composed a worldwide database for financial system benchmarking. In their paper, they argue the following:

”Better functioning financial systems stimulate new firm formation and help small, promising firms expand as a wider array of firms gain access to the financial system.”

This thesis focuses on the influence of the functioning of financial systems on en-trepreneurial activity using a self-constructed dataset. More concrete, this thesis ex-amines the relation between competitiveness and entrepreneurial activity in a multi-country setting. GMM panel data methods are used to empirically assess this rela-tionship and analyze the effects of bank market power on entrepreneurship. This has not been performed in earlier research1.

The remainder of this study is organized as follows. The second section gives a

(6)

literature review on research done in the field of financial systems and entrepreneurial activity. Section 3 presents the research method, where the data and models used in this study will be described. The fourth section presents the results and analysis. The fifth section contains a summary and the conclusions. Section 6 presents a discussion, where limitations and possible further research will be discussed.

(7)

2 Literature review

Entrepreneurship and its importance for an economy is widely recognized and much research elaborates on questions related to new firm formation. A distinction is made between research focused on practical issues as data gathering and measurement, re-search that examines the determinants of entrepreneurial activity and rere-search that investigates the relation between firm formation and economic output.

This section gives an introduction on both the existing theories regarding en-trepreneurial activity in relation to economic growth and determinants of firm forma-tion. Next to this, bank market structure and the accessibility of credit is examined. In addition, recent empirical studies on the topic will be discussed.

2.1 Entrepreneurial activity and economic growth

In his early work, Schumpeter (1934) came up with the concept of the pioneering entrepreneur as the main force behind economic development. The idea of creative destruction caused by innovations was new in Schumpeter’s view and he stated that this was vital for economic growth. He underlined the importance of bank credit by assuming it to be a first condition for new firm formation and hence innovation. The theory of Schumpeter is the first to link economic development to new enterprise foundation. His theory predicts that an increase in entrepreneurial activity fosters economic growth.

Nevertheless, this prediction was not empirically tested for a long time due to problems regarding the measurement and quantification of entrepreneurial activity (Wong et al., 2005). Therefore, the role of entrepreneurs was mainly elaborated and emphasized on in qualitative and descriptive research. Porter (1990) is one of the first to recognize entrepreneurial activity as the driver behind economic development of nations. Later, a study carried out by Wennekers & Thurik (1999) tries to synthesize different views linking entrepreneurship and economic growth. The authors conclude that entrepreneurship is a key factor in economic growth due to the globalization and hence the need for reallocation of resources. Wennekers & Thurik (1999) regard their study as a start in the field of research on entrepreneurship and economic development. Since this view of entrepreneurship as key factor within economic development is generally acknowledged, the focus of policy agendas of many governments moved to-wards enhancing the business climate, with a distinct design across countries (Van Stel et al., 2007). This increased attention for new policies on strengthening the business climate also engendered an increased amount of research activity in the field of en-trepreneurship and its determinants. This thesis tries to contribute on this topic, since

(8)

multi-country evidence is still scarce.

2.2 Determinants of new firm formation

Many studies have elaborated on the determinants of new firm formation. This is done from various perspectives such as the study of Armington & Acs (2002) examining the role of human capital and education or the study of Freytag & Thurik (2007) investi-gating country specific cultural variables. Verheul et al. (2002) suggest that analyzing the differences between countries regarding entrepreneurial activity implies that also factors other than economic ones are relevant. The authors try to synthesize existing literature in their so called Eclectic Theory. They state that entrepreneurship has an interdisciplinary nature and that analyzing it on one level is not satisfactory since the concept has to do with temporal, geographical, cultural, industrial and individual fac-tors. Their Eclectic Theory integrates different aspects and combines everything into a unified framework. Within this framework entrepreneurship is explained using a dis-tinction between the supply side and demand side of the concept as proposed earlier by Bosma et al. (1999). The supply side refers to labor market aspects, where character-istics of the population play an important role. The demand side of entrepreneurship refers to product market aspects, where economic and technological development are main indicators.

Next to the supply and demand variables, Verheul et al. (2002) introduce four other elements of entrepreneurship which together complete their widely acknowledged framework. These elements are institutions and government policy, individual deci-sion making, actual and equilibrium rates of entrepreneurship and cultural factors. In addition, the authors distinguish between the macro, meso and micro level of en-trepreneurship, where the national economy, industries and individual firms are studied respectively.

Many authors have built on this framework, contributing to it by providing ei-ther theoretical or empirical evidence. Garc´ıa (2014) empirically investigated different indicators within the framework and found, after dealing with multicollinearity and endogeneity, that tertiary education has a significant effect on new firm formation. Van Stel et al. (2007) did the same and found that labor market legislation and the minimum capital requirement for starting a business have significant effects on en-trepreneurial activity.

This study uses a macro level perspective on entrepreneurship. A multi-country panel dataset is used to analyze country specific variables. Hence, individual decision making and actual and equilibrium rates of entrepreneurship are not object of study

(9)

here. However, government policy and cultural factors are of considerable relevance in this study next to the more obvious demand and supply side factors influencing entrepreneurship.

2.3 Financing and concentration in the credit market

The traditional views of Schumpeter (1934) and King & Levine (1993) suggest that better financial systems engender more innovation and hence foster the entrepreneurial climate. In addition, the authors find that a well-developed financial system increases productivity and economic development. The financing of new businesses is a funda-mental aspect when analyzing new firm formation since financial capital is an essential requirement for firms to start up (Cassar, 2004). De Bettignies & Brander (2007) underscore that entrepreneurs prefer bank loans over other financing options due to the fact that it avoids loss of full ownership of the firm. Next to this, Huyghebaert & Van de Gucht (2007) suggest that banks may profit of having long-term connec-tions with promising new businesses. Bank financing for start-ups is proven to be of considerable importance for the foundation and success of new businesses (Robb & Robinson, 2014). Robb & Robinson (2014) emphasize that liquid credit markets are vital in this start-up process, because new firms depend on debt as start-up capital. Furthermore, Lemmon et al. (2008) show that when new firms grow their dependence on formal bank capital increases as well.

Concluding, it is widely acknowledged that bank funding plays an important role in new firm formation. This role is shaped by the organization and structure of the credit market in which the bank is operating. Hence, this makes the bank market structure in relation to new firm formation a debated subject of study (Agostino & Trivieri, 2016; Beck & Levine, 2002). Since the 1990s literature on the topic of bank concentration and competition in relation to access to bank credit has developed considerably (Berger et al., 2004). Economic theory distinguishes two distinct views about the sign of the effects. The traditional conception is referred to as the structure-performance hypothesis, where it is assumed that competitive banking markets foster access to credit (Beck et al., 2004). On the other hand, an alternative hypothesis is given by the information hypothesis where it is stated that informational asymmetries and agency costs lead to a weakened incentive of banks to issue loans (Carb´o-Valverde et al., 2009). This hypothesis suggests that less competition in the banking market relates to a greater accessibility of loans. Both views are examined by many researchers, either theoretically or empirically.

(10)

competitiveness in the credit market fosters or deteriorates financing of business start-ups. Guzman (2000) finds that more competitive credit markets lead to banks having a higher return on deposits and charging lower interest rates than banks in a monopolistic market. Hence, more credit market competition should foster new firm formation according to Guzman (2000), since a monopoly in the banking sector tends to lead to credit rationing. In contrast to the former, Petersen & Rajan (1995) suggest that new firms benefit from a credit market in which the concentration of banks is higher, as they are able to receive credit at better terms than in a more competitive credit market. Their reasoning builds upon the idea that banks expect to have long-term relationships with promising companies as borrowers.

Empirical research suggesting that there exists a positive relationship between bank competition and new firm formation has been done by, among others, Cetorelli & Strahan (2006) and Black & Strahan (2002). The latter used state-level US data to investigate whether more competition in the financial sector influences the growth rate of business creation. They link banking structures and competitiveness to en-trepreneurial activity and find that states with lower rates of new business formation have more concentrated local banking markets. Cetorelli & Strahan (2006) investigate the role of bank competition using data on local US markets for banking and non financial sectors. The authors examine whether a concentrated bank market is benefi-cial to the number of firms, firm size and firm size distribution within a market, using the following model:

yjst = α · xjst+ β · [bank dependencej· bank competitionst] + η (m) st + η

(i) jt + jst in this model the explanatory variable x is a vector of variables on employment share, market trends and industry trends for industry sector j in market s over time t. The model is re-estimated three times for different dependent variables; the number of firms, average firm size and a measure of the size distribution of firms. Using their panel dataset the authors construct an interaction term for the variable of interest, since the sheer effects of bank dependence and bank competition are captured by the two sets of fixed effects (market-level fixed effects η_st(m) and industry-level fixed effects η_jt(i)). The empirical findings of this article are in line with the notion that a concentrated banking market induces more obstacles for new firms to receive access to credit.

Furthermore, Beck et al. (2004) have assessed the influence of the structure of a nation’s banking market on the accessibility of companies to bank financing. They examine financing obstacles and find negative effects of more concentrated banking markets on access to credit for firms of all sizes, with a declining effect from small

(11)

to large firms. This result is in particular observable for developing countries. Beck et al. (2004) stress the importance of taking into account both the institutional and regulatory frameworks and the ownership structure of the banking market, as these influence a firm’s financing obstacles. The authors compose the following model:

Financing obstacle_it= α + β1· xit+ β2· concentrationit+ it

where xit denotes the vector of control variables for country i in year t. The coefficient of interest is β2. The model is estimated using an ordered probit model, since the dependent variable is polychotomous with a natural order. They find a significant positive coefficient which indicates that countries with more concentrated banking markets have higher financing obstacles (Beck et al., 2004). A similar conclusion using a different model is found by Cetorelli (2003), where the results indicate that a lower level of bank concentration is beneficial to the entry and growth of new firms. The findings are in line with the idea that higher banking market power implies more financial obstacles for potential entrants.

In contrast to the previous and as stated before, contrary findings have been found in earlier empirical research. DeYoung et al. (1999) use panel data of commercial banks in the US. The authors make a distinction between rural and urban markets and find that a higher level of concentration in the banking market positively influences small business lending in urban markets. However, DeYoung et al. (1999) find a slightly negative effect in rural markets. In addition, Petersen & Rajan (1995) examine US data on young firms and they find that banks and their lenders build a stronger relationship with each other in a highly concentrated banking market. Hence, their empirical results indicate that in US local banking markets with a high concentration rate, younger firms have a better chance of receiving bank financing. Corresponding to these findings are the results of Rogers (2012) after investigation of US market data. Rogers (2012) finds that more competition in the banking sector leads to less firm creation.

This thesis tries to contribute to the debate by investigating the relation between bank market power and entrepreneurial activity in a cross-country setting, which makes this paper one of the first examining this relationship using advanced GMM panel data methods on country level2.

2.4 Previous empirical studies

This paper investigates the relation between bank market power and firm creation in a multi-country setting by using GMM panel data methods. To be able to construct a

2

(12)

comprehensive model on this topic, some important recently examined variables and models are thoroughly discussed in this section.

2.4.1 Measuring entrepreneurial activity

Wennekers et al. (2005) make a distinction between a static and dynamic perspective on entrepreneurship, where the dynamic view indicates changes in the business ownership rate and the static just describes the number of business owners. These are also referred to as stock or flow indicators respectively (Garc´ıa, 2014). Flow indicators are for example indicators of nascent enterprise activities and birth of firms, whereas stock indicators are often given by indicators on business ownership and self-employment.

Several authors used static indicators such as self-employment for measuring en-trepreneurship (Van Praag & Versloot, 2007). However, this rate of self-employment might be biased due to the fact that not all self-employed are also entrepreneurs (Parker, 2004). In addition, the static indicator might by inaccurate because when entrepreneurs hire more workers the self-employment rate will decrease and there is no distinction between ownership of small and large firms. Hence, Glaeser (2007) insists on using a more dynamic perspective.

This paper measures entrepreneurship using a dynamic indicator. This is the num-ber of newly registered limited liability companies in a country, as was proposed by Garc´ıa (2014) and Agostino & Trivieri (2016). However, dynamic indicators do have disadvantages too (Garc´ıa, 2014), like the fact that the survival rate of new businesses is not incorporated in the business creation rate. Nevertheless, this study does not in-corporate research about the success of new businesses and hence a dynamic measure is sufficient.

2.4.2 Measuring bank market power

Traditionally, studies on competitiveness use concentration measures as proxy for bank market power. These concentration measures were seen as exogenous variables reflect-ing market power and as inverse indicators of competitiveness (Berger et al., 2004). Commonly known concentration measures are the Herfindahl-Hirschman Index (HHI) and the share of assets held by the top three or five largest banks. The HHI refers to the sum of the squared market share of each bank in the system (World Bank, 2015). However, empirical studies have shown that concentration measures, and the HHI in particular, as proxies for market power possess consistency and robustness problems (Hannan, 1997; Rhoades, 1995). In addition, the World Bank (2015) argues that con-centration does not measure competition properly and that direct indicators of bank

(13)

pricing behavior and market power are preferred.

Subsequently, many authors introduced alternative measures for competitiveness (Berger et al., 2004). These alternative measures include indicators on efficiency, qual-ity, risk and general economic involvement. Next to this, research on the topic ex-panded to a more dynamic and broader setting (Berger et al., 2004). The World Bank (2015) introduced three alternative measures: the Lerner index, the H-statistic and the Boone indicator. The Lerner index measures market power directly by calculating the difference between output prices and marginal costs as a fraction of prices (Carb´ o-Valverde et al., 2009). The H-statistic indicates the elasticity of bank revenues relative to input prices and is regarded as a concentration measure. The Boone indicator cap-tures the effect of efficiency on profits and is calculated by the elasticity of profits to marginal costs (World Bank, 2015).

Previous literature on industry level, both theoretical and empirical, examined the three indicators as estimators of market power. It is stated that the Lerner index is a structural measure of market power and a better measure than pure concentration measures (Connor & Peterson, 1992; Borenstein & Bushnell, 1999). As a consequence, Carbó-Valverde et al. (2009) investigated both the HHI and the Lerner index in the context of firm financing and bank market power. Carbó-Valverde et al. (2009) state that previous studies using solely concentration measures might give incomplete re-sults about the relationship between firm financing and bank market power. Therefore, Carbó-Valverde et al. (2009) compare both indicators and their results. The authors find theoretical evidence supporting that the Lerner index is a more comprehensive measure because it includes more factors that influence market power. However, op-posing empirical results are found by the authors when comparing the two measures statistically. By using the HHI, the authors find that more financing obstacles are present when bank market power is higher whereas the use of the Lerner index indi-cates lower financial constraints. Carbó-Valverde et al. (2009) argue that panel data constraints induce concerns about robustness problems, given these opposite results.

Since the use of both sets of measures remains a debated issue, this paper investi-gates both concentration measures and alternative measures and compares the results.

2.4.3 A model on firm financing and bank market power

Next to their investigation of concentration measures and alternative measures as proxy for bank market power, Carb´o-Valverde et al. (2009) empirically examined the relation between financing constraints and bank market power using data on Spanish small and medium-sized enterprises. They used panel data techniques to deal with endogeneity

(14)

problems that might arise from regressors about for instance efficiency and profitability of firms. The authors estimate the following model:

Bank loansit= β0+ β1· xit+ β2· [bank market power]_it+ it

To control for the endogeneity problems, Arellano-Bond GMM estimators are com-posed with a consistent variance-covariance matrix (Carb´o-Valverde et al., 2009). The explanatory variables are market power, using both a concentration measure and the Lerner index, and control variables, x, on firm, environmental and bank level. The dependent variable is a financing constraint variable. In the dynamic GMM case this variable is measured as the ratio of trade credit to tangible assets.

Carb´o-Valverde et al. (2009) find that the coefficients of the concentration measure and the Lerner index are both significant but with opposite signs. More concrete, the Lerner results indicate that a higher level of market power is accompanied with more financing constraints, whereas the concentration measure concludes the opposite. The authors further explored this inconsistency by also estimating the model using other concentration measures but this too did not give satisfactory results. They concluded that the inconsistency of these various measures leads to questioning the accuracy of this concentration measure as proxy for bank market power (Carb´o-Valverde et al., 2009). The main finding of their study is that the effects on the firm financing constraints of bank market power are sensitive to the choice for the bank market power proxy. This thesis will further investigate the differences between the various proxies by using both measures.

2.4.4 Models on bank market power and firm creation

Bonaccorsi di Patti & Dell’Ariccia (2004) are one of the first investigating the rela-tionship between new firm formation and bank market power using data on Italian industries. They state that competition is a precondition to new firm creation and estimate the following model using single country data:

Birth ratej = β0+ β1· g(bank market powerj) + β2· market characteristicsj + j

In this model the functional form of g(·) is incorporated in a linear and quadratic form. The authors find a bell-shaped relationship between bank market power and firm creation. However, these results might be considerably different than the results found in this study since the object of study are industries instead of countries.

The study of Agostino & Trivieri (2016) is one of the first addressing the relation-ship between new firm formation and bank market power in a multi-country setting.

(15)

This study will employ a similar dataset and will build forth on the model of Agostino & Trivieri (2016). A thorough discussion of their model will be carried out in subsection 3.2.1.

(16)

3 Methodology

3.1 Data

This subsection elaborates on the data used for this study. This thesis uses a self-constructed dataset, originating from various sources. The dataset is based on data and variables used in previous literature examining similar topics and to a consid-erable extent following the data used by Agostino & Trivieri (2016). The dataset consists of various main elements: three proxies for bank market power, a proxy for entrepreneurial activity and a wide range of control variables which control for eco-nomic, governmental and trend factors. The following section motivates the dependent variable and the explanatory variable of interest, which are entrepreneurial activity and bank market power respectively. Subsequently the set of control variables is introduced. This is followed by a subsection about data scarceness and how this thesis copes with this. To end the data section, some interesting summary statistics and suggestions for subsets are presented.

3.1.1 Main variables

As stated in the literature review, this thesis uses the new business entry density as a dynamic measure of entrepreneurial activity. The new business entry density is mea-sured as the number of newly registered limited liability corporations per calendar year, normalized by the working age population (new registrations per 1,000 people aged 15-64). This measure allows for comparing across heterogeneous legal regimes and economic systems (Klapper & Love, 2011). Furthermore, this definition is a stan-dard unit of measurement and takes into account data sparseness, relevance as measure of entrepreneurial activity, emphasis on the formal sector and consistency for differ-ent countries. Using only limited liability private companies is done because these companies are separate legal entities and it is the most common business form in the world (World Bank, 2009). The new business entry density includes all limited liability companies regardless of size. Partnerships and sole proprietorships are not considered because of different definitions and regulations around the world.

As previously mentioned, this study uses both a concentration measure as well as alternative measures as proxies for bank market power. The concentration measure used in this study is bank concentration measured as the assets of the three largest commercial banks as a share of total commercial banking assets. Hence, competi-tion within the banking sector is negatively related to the concentracompeti-tion measure. As claimed before, the predictive accuracy of this concentration measure on banking

(17)

com-petition is much debated on. Two alternative measures are the Lerner index and the Boone indicator, which are object of study in many recent studies (World Bank, 2015). These measures will also be considered in this study.

The Lerner index is based on markups in banking and is a measure of market power in the banking market. It compares output pricing and marginal costs (a markup) as a fraction of prices: L = (P − C0)/P . Marginal costs are obtained from an esti-mated translog cost function with respect to output and prices are calculated as total bank revenue over assets. An increase in the Lerner index indicates a deterioration in competitiveness within the banking system (World Bank, 2015).

The Boone indicator is a measure of degree of competition based on profit-efficiency in the banking market. It is calculated as the elasticity of profits to marginal costs. The elasticity is calculated by regressing profits on a measure of marginal costs. The resulting coefficient is the used elasticity. The idea behind the Boone indicator is that efficient banks yield higher profits. Higher values of the Boone indicator signal less competition in the banking market. The Boone indicator is included in the dataset in a transformed form, which is done in order to have a positive measure that increases when the degree of competitiveness decreases3.

Data for all three bank market power proxies is obtained from the Global Financial Development Database of the World bank, covering 206 countries for a period from 1997 to 20144_{. The data on entrepreneurial activity is obtained form the World Bank} Doing Business Database, covering 137 countries for a period from 2004 to 20145. Appendix A gives a description of all included variables and their sources.

3.1.2 Control variables

Following the framework by Verheul et al. (2002), control variables are segregated into various categories. These categories are demand side, supply side and government regulations, all determinants of the new business entry rate. Cultural factors are assumed to be time invariant and hence are captured by the unobserved country specific time invariant fixed effects6.

As for demand side control variables, two variables are included. Firstly, annual

3_{The transformation of the Boone indicator proceeds as follows. First, the minimum and maximum}

value of the absolute original values are calculated. Subsequently, the difference between each original value and the sum of this minimum and maximum value is calculated. In the final dataset, these differences are considered in absolute form

4_{The data and sources can be found on:}

http://data.worldbank.org/data-catalog/global-financial-development

5

The data and sources can be found on: http://econ.worldbank.org/research/entrepreneurship

6

(18)

GDP variation is taken into account, which is measured as an annual percentage of GDP growth. Verheul et al. (2002) state that the effect of economic growth on en-trepreneurial activity is ambiguous, depending on the stage of economic development of a country7. It is expected that the influence of GDP growth on new business entry density is positive but different across countries with different income levels. To deal with the distinction between levels of economic development an interaction term will be tested. A second examined demand side indicator is foreign direct investment net inflows as percentage of GDP. Both indicators are extracted from the World Develop-ment Indicators database of the World Bank8.

As for supply side determinants of entrepreneurship, a distinction is made between more general supply side determinants and banking supply side determinants. For the general case, three variables are included within this study. First, population growth is expected to have a positive influence on new firm formation. Verheul et al. (2002) find that countries with a growing population, and hence work force, show a positive impact on the share of entrepreneurs. Secondly, attention is paid to the female labor force participation rate. This is included in the dataset as the proportion of the female population aged 15-64 which is economically active. Verheul et al. (2002) argue the importance of this measure since the share of female business owners is increasing. It is assumed that this has a positive influence on new firm formation. Finally, educa-tion rates are assumed to have a positive relaeduca-tion with new firm formaeduca-tion (Agostino & Trivieri, 2016). Here, the education rate is measured as the total enrollment in secondary education expressed as a percentage of the population of official secondary education age. These three indicators are extracted from the World Development Indicators database of the World Bank.

Three banking determinants are included in the constructed dataset. First, private credit by deposit money banks and other financial institutions as percentage of GDP is included. This measure is included since it incorporates the financial resources provided to the private sector by banks and other financial institutions and this plays an important role in the start-up process of businesses. Next to this, the aggregate bank Z-score is included in the dataset. The Z-score captures the probability of default of the commercial banking system in a country and compares the buffer of the commercial banking system with the volatility of returns (World Bank, 2015). Both measures are assumed to have a positive relation with new firm formation (Agostino & Trivieri, 2016). Finally, a banking crisis dummy is included, which indicates 1 for the presence

7

For a thorough discussion of the literature on this subject, see the article of Verheul et al. (2002).

8

The data and sources can be found on

(19)

of a banking crisis and 0 otherwise. Obviously, this dummy is expected to have a negative relation with the new business entry rate. All the banking indicators are obtained from the World Bank Global Financial Development Database.

The previous sections show the influence of market forces on both the demand and supply side level of entrepreneurship. Verheul et al. (2002) stress the importance of also controlling for government legislation regarding entrepreneurship. To address this, three variables are included originating from the Doing Business database of the World Bank. First, an indicator on the minimal capital requirement for starting a business is included. The variable used is the minimum paid-in capital required to start a business as percentage of income per capita. Minimum capital requirements are considerably higher in low income countries (World Bank, 2013) and are assumed to negatively in-fluence entrepreneurship. The second added variable is the cost of resolving insolvency, in other words the bankruptcy costs. The World Bank (2009) emphasizes the impor-tance of this measure within the business framework. It is presumed that there is a negative relation between this measure and new firm formation. Thirdly, an indicator on public registry coverage as percentage of all adults is included. Tore public credit registry coverage is expected to positively impact firm creation (Agostino & Trivieri, 2016).

3.1.3 Data scarceness

The data on the new business entry density rate, which is the dependent variable in this study, is available for 137 countries from 2004 until 2014. However, for the years 2013 and 2014 the data on this density rate is very scarce and hence these two years are omitted from the dataset. In addition, countries with a complete lack of data on the bank market power proxies are dropped. For the concentration measure and the Boone indicator this means that 20 more countries are omitted from the dataset. For the Lerner index 7 more countries are removed. Furthermore, countries lacking all government indicators and three or more subsequent data points for the business entry density rate are removed from the dataset. This leaves 93 countries for the concentration measure and Boone indicator and 86 countries for the Lerner index. The process of removing countries when examining variable availability and hence the number of included countries and years in the dataset are stated in Table 1. For the remaining gaps in the data of the control variables, linear interpolation is applied to fill in the gaps. This is a quite strict assumption but due to the fact that it is applied only at very few data points it is a good solution for minimizing data loss.

(20)

data sparseness. More developed countries often have more available data, since data gathering is more easy and registry systems are more advanced. The deleted countries possessing less data are therefore mostly lower middle or low income countries. This gives, obviously, a possibility for bias in the results. Therefore, the results can only be hold valid for the included countries. A list of all countries included can be found in Appendix B.

Table 1: Number of countries included in the dataset

Variable included # countries left in dataset

Business entry density rate 137

Bank market power proxies 117

Government indicators 115

Business entry density rate 3 subsequent gaps 93*

Lerner index 86**

* # countries for Concentration and Boone cases ** # countries for Lerner case

3.1.4 Summary statistics and subsets

Many previous studies have examined business entry rates and found heterogeneity in the rates between countries for different income levels (Klapper & Love, 2011). In ad-dition, Verheul et al. (2002) state that the effect of economic growth on entrepreneurial activity differs between the developing state of various countries. To give more insight on the dissimilarity, this section will give some summary statistics showing relevant differences per country income level.

The data about income groups is obtained from the World Bank list of economies 2016 (Fantom & Serajuddin, 2016), which classifies all countries into four income groups: low, lower middle, upper middle and high income using gross national income per capita9. Figure 1a shows the different levels of new business entry density rates per income group using 2004-2012 averages. This study incorporates two ways of examining the different effects per income group. First, interaction terms are tested containing dummy variables of various income groups and the bank market power proxy or GDP growth. Second, the results of the whole dataset will be compared to the results of relevant subsets of the data.

In addition, the World Bank list of economies divides the countries in seven regions: South Asia, Europe & Central Asia, Middle East & North Africa, East Asia & Pacific,

(21)

Sub-Saharan Africa, Latin America & Caribbean and North America10_{. To investigate} the geographic effects, subsets of relevant regions are estimated as well. Figure 1b shows the entry density for these seven regions, using 2004-2012 averages. Important to note is that within the countries included for this thesis the North American region only consists of Canada and hence is not a good reflection of the whole region.

As stated by Klapper & Love (2011), the business entry density rate is expected to gradually grow before the financial crisis and gradually decline after. To get more insight in this trend and to discover differences between income groups Figure 2 is constructed. Figure 2 shows this trend especially for high and upper middle income groups. To deal with the expected long term trends as shown by Klapper & Love (2011) a linear trend variable will be tested in the estimated model. This trend variable will also be tested using it within an interaction term with a dummy variable for income level, to account for the difference in trend for countries with various income levels.

Figure 1: Entry density levels for subsets, 2004-2012 averages (a) Entry density by income group

Low Low ermiddle Upp ermiddle High 0 2 4 0.29 0.58 3.52 5.01 Business en try densit y rate

(b) Entry density by region

South Asia North America Middle East&North Africa Sub-Saharan Africa Latin America&Caribb ean Europ e& Cen tral Asia East Asia &P acific 0 1 2 3 4 5 0.19 0.97 1.19 2.05 3.18 4.07 4.76 Business e n try densit y rate

Figure 2: Entry density trends, income group averages

2004 2006 2008 2010 2012 0 2 4 6 Year Business e n try densit y rate low lower middle upper middle high 10_{See appendix B}

(22)

3.2 Models & Methods

3.2.1 Reviewing the model of Agostino & Trivieri (2016)

Since this thesis builds forth on the results of Agostino & Trivieri (2016) (hereafter referred to as AT), their model will be analyzed in this section. AT estimate the following equation:

yit= β0+ β1bit+ β2b2it+ γxit+ δhi+ X

s

θsd(s)t + ηi+ it (1)

In this model, yitrepresents the new business entry density, bitrefers to a bank market power proxy which is the explanatory variable of interest, xit is a vector of control variables, hi is a dummy variable for high income countries, Psθsd(s)t gives a summa-tion over the time fixed effects, ηi represents the country fixed time invariant effects and it are idiosyncratic errors.

The exact dataset of AT is not (yet) publicly available. However, the author of this study tried to recompile the used dataset as precisely as possible, using the same World Bank sources. However, the set of countries included in the AT dataset differs from the recompiled dataset due to differences in data availability. Subsequently, the AT results are replicated for the new dataset and similar results are found (not reported). To reanalyze the model of AT, several tests on the functional form of (1) are performed. AT do not give a satisfactory motivation for specifying their model as it is and in this thesis the model is reviewed by testing it on the functional form and misspecification11.

To begin with testing an employed panel data model, one of the most logical things to review first is the decision for either random or fixed effects. This is done using the Hausman test with the null hypothesis that random effects should be estimated instead of fixed effects. If the Hausman test rejects it is assumed that the individual effects are significantly correlated with one or more regressors and hence random effects give problems (Park, 2011). Performing the Hausman test on model (1) rejects the null hypothesis (p-value = 0.0018) and hence fixed effects should be incorporated in the model. However, AT estimate model (1) with a random effects estimator, which should give problematic results when taking into account these results of the Hausman test.

In addition, AT do not give any reason for including the squared bank market power proxy in their model. To test whether a non-linear functional form is pre-ferred over a linear regression model, the Ramsey Regression Equation Specification Error Test (RESET) test can be executed. The RESET test indicates whether non-linear explanatory variables have power in explaining the dependent variable (Ramsey,

(23)

1969). Performing the test shows a rejection of the null hypothesis (p-value = 0.000) of no omitted variables so this indicates that indeed other interactions and squared explanatory variables are necessary to obtain a satisfactory model12. Investigating the explanatory power of squared terms can be done directly by adding them in the model and analyzing their significance.

AT examine a static model as given in (1). However, a dynamic model might be more appropriate and this can be tested using several tests on the autocorrelation within the model. However, to address the dynamics directly, it is more straightforward to just include lags of the dependent and explanatory variables as regressors and see whether these coefficients show any significance. Following the economic literature, it is expected that the model of AT suffers from neglected dynamics. To test this for the AT model, a new model is estimated where the first lag of the dependent variable is added to the model of AT13:

yit= αyit−1+ β0+ β1bit+ β2b2it+ γxit+ X

s

θsd(s)t + ηi+ it (2)

After obtaining the GMM panel data estimators as explained in subsection 3.2.3, it is found that the parameter estimate of the lagged dependent variable, α, is highly significant (p-value = 0.000). The static model assumes that the changes in explanatory variables only have influence in the same year. However, after a review of the model of AT, it is proposed that a dynamic model should be used instead of a static model, since the dynamics here cannot be neglected. Summarizing, reviewing the AT model gives suggestions for an improved model.

3.2.2 ADL models

As proposed in section 3.2.1, a dynamic model should be employed in this thesis. Suf-ficient lags of all explanatory variables should be included as regressors. A balance between including enough lags and not imposing too many restrictions should be found in order to have a proper model. The modeling framework used for this is the Autore-gressive Distributed Lag (ADL) framework. To be able to elaborate on the dynamics of this model, the fairly general ADL(1,1) model is considered:

yit= αyit−1+ β0xit+ β1xit−1+ it (3) i = 1, ..., N ; t = 1, ..., T

12

The Ramsey Reset test is formally only constructed for the simple OLS model, but for a panel data model with fixed effects its extension is straightforward

13

The model in (1) is estimated with random effects. Since model (2) is estimated using a

(24)

where yit denotes the new business entry rate, yit−1 denotes the new business entry rate in the previous year, xit is a vector of all current regressors, xit−1 is a vector of the one year lagged regressors and it is supposed to be white noise (Kiviet, 2017).

Classifying the model as an ADL(1,1) model has the advantage of easily being able to interpret the long run effects in equilibrium. To show this, the model in (3) can be rewritten using the lag operator L:

(1 − αL)yit= (β0+ β1L)xit+ it (4)

The model is stable when (1 − αL) has all roots outside the unit circle (Kiviet, 2017). This means that for stability, α should have an absolute value smaller than one14: |α| < 1.

To investigate the long run effects it is assumed, for simplicity, that xit is a scalar instead of a vector. The long run multiplier of y with respect to x is given by:

¯ yi= β0+ β1 1 − α0 ¯ xi (5)

The change in the dependent variable can be explained as the sum of two components, as shown by Keele & De Boef (2004). The author substitutes yit by yit−1+ ∆yit and xit by xit−1+ ∆xit in equation (3) and this can then be rewritten as:

∆yit= β0∆xit− (1 − α)[yit−1−

β0+ β1

1 − α xit−1] + it (6)

Hence, the first component, β0∆xit, gives the change proportional to the contempo-raneous change in x. The second component, −(1 − α)[yit−1− β_1−α0+β1xit−1], gives an adjustment of the equilibrium error, in other words a correction for the deviation from the equilibrium in the previous period. The dynamic impacts of the estimated model will be further discussed in subsection 3.3, after the employed model is determined.

3.2.3 Panel data techniques

This thesis employs panel data to examine the effect of new firm formation on bank market power. The increased precision in estimation is a major advantage of panel data methods over cross section methods. Furthermore, panel data can incorporate underlying microeconomic dynamics and allows for unobserved individual heterogene-ity (Cameron & Trivedi, 2005; Bond, 2002). In practice, it is proven to be difficult to find valid instruments which correct for the omitted variables bias originating from unobserved heterogeneity. Panel data offers an alternative by making use of internal instruments.

(25)

Strict exogeneity of regressors is not a plausible assumption in the case of economic models since it is likely that these models contain jointly dependent variables and present jointly determined results. For a panel data set containing a limited number of periods and large number of subjects and in which joint dependence may be present, Generalised Method of Moments (GMM) is a convenient and conventional technique. In general, the GMM method makes the following assumptions for dynamic models on the data-generating process (Roodman, 2006):

1. The number of periods, T, is small. The number of subjects, N, is relatively large.

2. The process might by dynamic

3. Arbitrarily distributed time invariant individual effects might be present

4. Idiosyncratic disturbances might have individual specific patterns of heteroskedas-ticity and no serial correlation

5. The idiosyncratic disturbances are independent across subjects

6. Some regressors might be endogenous or predetermined

Consider the following dynamic panel data model for which the above assumptions hold:

yit= x0itβ + wit0 γ + vit0 δ + τt+ ηi+ it (7)

where yitis the dependent variable, xitcontains strictly exogenous additional (possibly lagged) explanatory variables (xit is uncorrelated with all past, present and future realizations of it), witcontains predetermined additional (possibly lagged) explanatory variables (wit is uncorrelated with it, but might be correlated with it−1 and earlier, lags of the dependent variable are possibly included in wit), vit contains endogenous additional explanatory variables (vit is correlated with it and earlier shocks, but vit is uncorrelated with it+1 and subsequent shocks, hence jointly dependent with yit), τtare unobserved random or fixed individual invariant time effects, ηi are unobserved individual specific time invariant effects which contribute to the heterogeneity in the means of the yit series across individuals and it is the disturbance term which is assumed to be independent across subjects and serially uncorrelated. The presence of unobserved heterogeneity within the model has to be addressed. Also because the possibility of correlation between the explanatory variables (either xit, wit or vit) with the individual effects ηi can not be neglected. Panel data is a solution since

(26)

it has the potential to still obtain consistent parameter estimators by performing a transformation to eliminate individual effects.

Arellano & Bond (1991) introduce dynamic panel estimators which are commonly used and popular in practice. Arellano-Bond start with transforming the regression and then applying GMM techniques. Their transformation usually means first differencing and is therefore also referred to as difference GMM. For (7), first differencing results in

∆yit = ∆x0itβ + ∆w0itγ + ∆v0itδ + ∆τt+ ∆it (8)

Taking these differences introduces an endogeneity problem since a correlation be-tween a regressor, for example ∆yit−1 which is assumed to be included in wit, and the error term, ∆it, emerges through it−1. To estimate the coefficients and to deal with this endogeneity, internal instruments are introduced. These instruments contain lags of explanatory variables which meet the orthogonality conditions expressed as the following GMM moment conditions:

• E[yi,t−j∆it] = 0, j ≥ 2; t = 3, ..., T

• E[x0i,j∆it] = 0, j = 1, ..., T ; t = 3, ..., T , for xit exogenous with respect to it. • E[wi,t−j0 ∆it] = 0, j ≥ 1; t = 3, ..., T , for wit predetermined with respect to it. • E[v0

i,t−j∆it] = 0, j ≥ 2; t = 3, ..., T , for vit endogenous with respect to it. The transformed error term is uncorrelated with the above specified lags (and con-temporaneous variables in the exogenous case) of the explanatory variables and this gives a suggestion for appropriate internal instruments. Hence, these orthogonality conditions imply an instrument matrix Z for which the following moment condition holds:

E[Zit0 ∆it] = 0, t = 3, ..., T (9)

As in Kiviet et al. 2017, rewriting and stacking the time-series observations gives:

∆yit= ∆rit0 α + ∆˜ it (10)

˜

yit= ˜R0iα + ˜˜ it (11)

where ˜R0_i comprises all the transformed explanatory variables.

Let G be a weighting matrix of which the form depends on using one-step or two-step estimators. In short, the two-two-step weighting matrix uses the consistent one-two-step

(27)

residuals to obtain an asymptotically optimal weighting matrix (Kiviet et al., 2017). The GMM estimator for ˜α yields

ˆ ˜

αGM M = ( ˜R0ZGZ0R)˜ −1R˜0ZGZ0y˜ (12)

One-step GMM estimators rather than two-step estimators are often explored in em-pirical literature using Arellano-Bond, since simulation studies have shown the limited improvement in both bias and variance of the two-step estimator even in the case of notable heteroskedasticity (Bond, 2002; Kiviet et al., 2017).

Blundell-Bond amplify the assumptions of Arellano-Bond by assuming that first differences of lagged regressors are uncorrelated with the fixed effects (Roodman, 2006). Hence, more instruments can be introduced next to only level instruments with Arellano-Bond. This may improve efficiency considerably. Blundell-Bond esti-mators are also referred to as system GMM since two equations are used: both the differenced equation as well as the level equation (Arellano & Bover, 1995). The valid first differenced instruments for the Blundell-Bond case are used to instrument in the level equations. These can be estimated under effect stationarity of all variables (Kiviet et al., 2017), which yields

E[ritηi] is time-invariant (13)

so that E[∆ritηi] = 0, ∀i, t = 2, ..., T (14)

where rit = (x0itw 0 itv

0 it)

0_. _{The extra imposed orthogonality conditions for the level} equation imply that the following must hold

• E[∆x0i,j(ηi+ it)] = 0, j = 1, ..., T ; t = 3, ..., T , for xit exogenous with respect to it.

• E[∆w0

i,t−j(ηi+ it)] = 0, j ≥ 1; t = 3, ..., T , for wit predetermined with respect to it.

• E[∆vi,t−j0 (ηi+ it)] = 0, j ≥ 2; t = 3, ..., T , for vit endogenous with respect to it. These added moment conditions imply that yi,t−1is instrumented with ∆yi,t−1. Hence, the orthogonality of the instruments relative to the error term of the level equation, (ηi + it), is not obvious. This follows from the fact that yit−10 contains the fixed effect ηi and the error term of the level equation also contains ηi. Summarizing, effect stationarity implies that the relation between the regressors and the fixed effects is constant over time, such that the differenced instruments are valid and the added orthogonality conditions hold (Blundell & Bond, 1998)15. As stated by Roodman

(28)

(2009), when the value of the coefficient of the lagged dependent variable is close to one, the effect stationarity condition is likely to be violated.

Two-step GMM is preferred when obtaining Blundell-Bond estimators since there is no known weighting matrix which results in an one-step GMM estimator which is asymptotically equivalent to the optimal two-step GMM estimator.

Summarizing, Arellano-Bond instruments the differenced equation with level in-struments. Blundell-Bond augment this by also using differenced instruments for the level equations. The Blundell-Bond estimator is preferred, if the extra moment condi-tions hold i.e. under effect stationarity, since it improves efficiency when taking these additional conditions into account.

3.2.4 Specification tests

Several specification tests are available for the Arellano-Bond and Blundell-Bond GMM estimators. The Hansen J-test testing the overidentification restrictions is always pre-sented with the GMM results. Overidentification is often the case and the J statistic is χ2 distributed with the degrees of freedom equal to the level of overidentification. The test is used for testing the joint validity of the instrument set, in other words whether the orthogonality conditions are satisfied. It is commonly interpreted as a test for validity of specification (Roodman, 2009).

In addition, a test for the validity of subsets comes along with the GMM estimators. This is referred to as the difference-in-Hansen test and the test statistic is again χ2 distributed with degrees of freedom equal to the number of instruments tested. The test gives a J statistic based on the increase when the considered subset is added to the instrument set, assuming that the other instruments are valid. In the Blundell-Bond situation special attention should be given to the difference-in-Hansen test for the differenced instruments in the level equations. As mentioned before, the assumption that the lagged differenced instruments are uncorrelated with the error term is not trivial (Roodman, 2006).

Furthermore, the GMM results are accompanied with serial correlation tests for the error terms to give insight in the reliability of the results. The first-order autocorrela-tion test should reject the null hypothesis of no serial correlaautocorrela-tion since by construcautocorrela-tion this should be present. Second-order autocorrelation should not reject the null hypoth-esis since otherwise this might indicate omitted variables in the model. Moreover, as proposed by Bun & Windmeijer (2010) the variances of the unobserved heterogeneity and the idiosyncratic errors will be compared to determine if the employed model is powerful. This is the case when the two variances do not differ considerably.

(29)

Generally, standard errors of two-step GMM tend to be downwards biased. Arel-lano & Bond (1991) aimed to improve finite sample performance in their study by introducing standard error corrections. For the one-step GMM heteroskedasticity ro-bust standard errors can be obtained when using Arellano Bond en Blundell Bond estimators. Similarly, for two-step GMM Windmeijer corrected standard errors can be requested (Windmeijer, 2005). The robustified and corrected standard errors will be used in this study, as is common practice in recent panel data studies and is recom-mended by, among others, Roodman (2006).

To finish this section of GMM techniques and specification, a note must be made about the size of the instrument set. In the case of a limited number of subjects, reducing the instrument count is often helpful. Two main techniques are commonly used to reduce the number of instruments (Roodman, 2009). Firstly, it is possible to collapse the instrument set. This reduces the number of moment conditions which can be useful when the number of subjects in the data set is relatively sparse. Furthermore, size problems of the overall Sargan-Hansen overidentification tests and the incremental Sargan-Hansen effect stationarity test can be moderated by collapsing the instrument set (Kiviet et al., 2017). The second technique for reducing the instrument count is limiting the number of employed lags used as instruments.

Summarizing, GMM panel data methods are accompanied with making a fair num-ber of choices on different aspects. As motivated before, this thesis will present het-eroscedasticity robust one-step Arellano-Bond estimators and Windmeijer corrected two-step Blundell bond estimators, provided that the implied moment conditions hold. The aforementioned assumptions and specification tests will be taken into account when constructing the model and results will be presented in section 4 after specifying the estimated model in the next section.

3.2.5 Estimated model

The previous sections have explained the proposed form of the model and the proposed techniques, namely an ADL model framework using Arellano Bond and Blundell Bond estimators. This section will elaborate on the specifics of the regression equations. Before doing so, it must be noted that the whole procedure of determining the right regression equation is a dynamic process. Furthermore, using panel data models comes along with finding a good balance between generality of the specification and imposing not to many untested restrictions.

Before starting with the practical part of construction of the model, some theoret-ical notes should be made. First, the necessary condition of identification indicates

(30)

that the number of instrumental variables must be higher than the number of unknown explanatory variables. Moreover, the instrumental variables should be uncorrelated with the error term. As often in econometric theory, asymptotic consistency and effi-ciency play an important role in constructing a satisfactory model. Valid instruments contribute to consistency of the model and effective instruments make the estimates ef-ficient (Kiviet, 2017). Next to this, dynamics of the assumed relation can be captured by including lags of explanatory variables and the dependent variable. Taking into account the requirement of no serial correlation of the disturbances, sufficient relevant lags should be included.

A model which includes relevant explanatory variables but does not impose too many moment conditions is constructed. Finding the most adequate functional form is done by testing various squared terms and interaction terms16. Because the rela-tionship is assumed to by dynamic, a number of lags of the dependent variable and the explanatory variables are included and tested on explanatory power. Highly in-significant variables are omitted since they have no predictive power. After thoroughly pretesting and refining, the resulting model which has the most empirical support is given by:

yi,t = αyi,t−1+ β0bi,t+ β1bi,t−1+ γ0xi,t+ γ1xi,t−1+ δ0zi,t+ δ1t + τt+ ηi+ i,t (15)

In this model, yi,t represents the new business entry density rate for country i in year t and this depends on the entry rate in the previous year yi,t−1, bank market power in the current and previous year, bi,t and bi,t−1, a vector of contemporaneous control variables xi,t and one year lagged control variables xi,t−1, a vector of interaction terms zi,t, a linear trend variable t, time specific effects τt, country specific effects ηi and the idiosyncratic errors i,t.

As stated before, three different proxies are used for the bank market power vari-able, namely the Lerner index, a concentration measure and the Boone indicator. The set of control variables consists of17:

• Demand side variables: GDP growth and FDI net inflows as percentage of GDP. Both included in the contemporaneous and lagged vector.

• General supply side variables: population growth, secondary education enroll-ment and female labor participation. Both included in the contemporaneous and lagged vector.

16

This is done in a structured way, by testing all possible relevant interaction terms and analyzing the various tests and significance levels

(31)

• Banking supply side variables: credit availability, banking crisis dummy and banking z-score. All included in the contemporaneous set. Only credit availabil-ity is included in the lagged vector, since the crisis dummy does not variate much over time and both lags do not seem to have much explanatory power.

• Government variables: starting a business costs, closing a business costs and public registry coverage. All three only included in the contemporaneous vector, since government regulations and interventions need time to adjust and do not variate much over time. The lagged variables did not seem to have much ex-planatory power since their coefficient is close to zero and the significance level is low.

Adding interaction terms to the model is accompanied with making choices about which terms to include and which not to include. After pretesting and analyzing the economic literature and summary statistics as mentioned in previous sections, the vector of interaction terms includes the following relevant interaction terms:

• Interaction term 1: linear trend multiplied by a dummy variable for low income countries.

• Interaction term 2: GDP growth multiplied by a dummy variable for low income countries.

• Interaction term 3: bank market power multiplied by a dummy variable for high income countries.

Interaction terms including a variable on the different regions (see section 3.1.4) are not included since they give highly insignificant results. Hence, to still be able to make a distinction between regions, different subsets are tested in subsection 4.4. The following section will elaborate on the specifics of the GMM regression and chosen instrument set.

3.2.6 Constructing the instrument set

The model in (15) will be estimated using the Arellano-Bond heteroskedasticity ro-bust one-step estimator and Blundell-Bond Windmeijer corrected two-step estimator, as proposed in section 3.2.3. As mentioned before, it is important to not include too many instruments as the number of subjects, countries, is relatively small. To avoid this problem, only internal instruments up to the third lag will be employed and the instrument sets will be collapsed. Further lags are assumed not to be very powerful

(32)

instruments. In addition, Kiviet et al. (2017) state that lowering the number of instru-ments by dropping weak instruinstru-ments might improve the precision of the results since problems with the bias in finite samples are reduced.

To be able to perform Arellano-Bond and Blundell-Bond techniques the explana-tory variables should be divided in three subsets of exogenous, predetermined and endogenous regressors. The nature of the variables determines the employed lags for the instrument sets. The classification is done using the results of the incremental Hansen test. First, all unlagged regressors are assumed to be endogenous since this is the weakest assumption. Sequentially all subsets of instruments are tested if they are also valid if the regressors were actually predetermined by including the first lag in the instrument set. If the incremental Hansen test does not reject, the variable is classified as predetermined. For these variables the first lag is included as instrument. To avoid influence of the order of testing, the variable with the highest p-value is treated as predetermined first. This is repeated until all variables are either classified as predetermined or endogenous. Similarly, all unlagged variables now classified as predetermined are tested on being exogenous. This is executed as before, but now including sequentially the contemporaneous variables in the instrument set. By defini-tion, the year dummies and the trend variable are both exogenous. The classification results of the Lerner case are shown in Table 2. Results of the Concentration and Boone cases are given in Appendix C.

3.3 Dynamic Impacts

This section will elaborate on the dynamic impacts which originates from including lagged regressors. As briefly shown in section 3.2.2., ADL models make it fairly easy to interpret long term effects in equilibrium, where a distinction can easily be made between contemporaneous and past effects. For this section, the ADL model in (15) will be used to determine the dynamics.

The explanatory variable of interest in the estimated model is bank market power, which appears in the model as contemporaneous variable, first lag and in an interaction term with a dummy variable for a country’s income level. To gain understanding of the dynamics with respect to bank market power, the model in (15) is rewritten as:

yi,t = αyi,t−1+ β0bi,t+ β1bi,t−1+ δ0[bi,t∗ ci] + γ0xi,t+ γ1xi,t−1+ τt+ ηi+ i,t (16)

Where it is assumed that |α| < 1 holds, as showed in section 3.2.2. For simplicity of computation, the other interaction terms and linear trend variable are left out here since they do not affect the dynamics with respect to bank market power. ci represents

(33)

Table 2: Classification of variables

Variable p-value Hansen p-value Hansen Classification predetermined* exogenous**

Lerner index 0.870 0.930 Exogenous

GDP Growth 0.842 0.677 Exogenous

FDI 0.124 NA Endogenous

Population Growth 0.004 NA Endogenous

Female labour participation 0.83 0.235 Predetermined

Education 0.200 NA Endogenous

Credit 0.894 0.634 Exogenous

Banking crisis dummy 0.906 0.302 Predetermined

Bank Z-score 0.033 NA Endogenous

Starting a business: costs 0.144 NA Endogenous

Closing a business: costs 0.684 0.654 Exogenous

Registry coverage 0.773 0.563 Exogenous

Interaction term 1 0.949 1.000 Exogenous

Interaction term 2 0.049 NA Endogenous

Interaction term 3 0.834 0.719 Exogenous

* Assuming endogeneity and testing for predeterminedness ** Assuming predeterminedness and testing for exogeneity

a dummy variable for low income countries (1 if country is classified as low income country, 0 otherwise). Successive substitution yields18:

yi,t= β0bi,t+ ∞ X l=1 αl−1(αβ0+ β1)bi,t−l+ δ0 ∞ X l=0 αl[bi,t−l∗ ci]+ γ0xi,t+ ∞ X l=1 αl−1(αγ0+ γ1)xi,t−l+ (1 − α)−1ηi+ ∞ X l=0 αli,t−l (17)

Since ci is time invariant, (17) can also be written as:

yi,t = [β0+ δ0ci]bi,t+ β1bi,t−1+ αyi,t−1+ ηi+ i,t−l (18)

Hence, the immediate impact multiplier for bi,t is β0 + δ0ci. The total multiplier can be computed and looks similar to the simple ADL(1,1) total multiplier as given in (5). In the case of model (15), the total multiplier, the effect on yi,t of a permanent one unit change in bi,t, is given by:

T Mb = (β0+ β1+ δ0ci)(1 − α)−1 (19)

(34)

The dynamic impacts are constructed for the various results, where both immediate as well as long run effects will be calculated. However, to be able to compare the results of the different bank market power proxies, some further calculations should be made. This will be elaborated on in subsection 4.3. The expectation is that a more competitive banking market has a positive immediate effect and a larger positive long run effect on entrepreneurial activity. As stated before, a stable dynamic process requires a coefficient of the lagged dependent variable smaller than one in absolute value.

Linking competitiveness in the banking sector with entrepreneurial activity : Evidence from dynamic panel data methods in a multi-country setting