The effect of start-up firms on unemployment : in 24 European countries from 2008-2012

(1)

The effect of start-up firms

on unemployment

In 24 European countries from 2008-2012

Cato Kleipool

10356185

BSc Economics & Finance

June 24, 2015

(2)

Statement of originality

This document is written by Student Cato Kleipool 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.

(3)

Introduction

Something of a controversy has arisen in literature about the relationship between the birth of new firms and unemployment. Two different views have emerged on this complex relationship. One view, which is referred to as the ‘unemployment push’, or ‘refugee’ effect, suggests that unemployment and new firm creation are positively related. Studies on self-employment (e.g. Evans & Leighton, 1989) have found that the unemployed are more likely to become self-employed, so higher levels of unemployment increase the number of startup firms. Another view, which has been called the ‘unemployment pull’, or ‘entrepreneurial’ effect, turns the causal relationship around. It suggests a negative relationship and poses that new firms and entrepreneurship contribute to reduced levels of unemployment by creating new jobs.

The purpose of this paper is to provide a study on the latter relationship. Studies that include both effects find that the unemployment pull effect is stronger than unemployment push effect. Therefore this paper will look only at the unemployment pull effect. It will attempt to answer if and how the startup rate has influenced unemployment in 24 European countries since the onset of the crisis in 2008.

The nature of the relationship between the number of new firms and unemployment is of great importance in the context of public policy, as policy makers strive for lower levels of unemployment. Understanding the true relationship between entrepreneurship and unemployment can guide policy makers when deciding if and how to promote entrepreneurial activity.

To answer the question whether new firm births have influenced unemployment across European countries since 2008 this paper will extend an existing model on unemployment and its causes. In their influential paper Nickell, Nunziata and Ochel (2005) explain a large part of unemployment through labor market institutions such as the level of unemployment benefits, trade union power, income taxes and the degree of wage flexibility. In this paper, the startup rate is added to their regression to find if this significantly increases the part of unemployment that can be explained through the model. Nickell et al. perform a panel data analysis. Data on the number of firm births and deaths is available for most European countries from 2008 and onwards only. Because labor market institutions and the number of startups will most likely not have changed by much in such a small period of time, this paper will perform a cross-sectional analysis and only look at differences between European countries, rather than a panel data analysis.

In the first section the available literature on the subject is reviewed and the baseline model by Nickell, Nunziata and Ochel is explained. Section two lays out the econometric analysis with information on the model, hypotheses, data and results. No significant influence of startups on unemployment is found. In section three a discussion is presented on the model and its results and suggestions are made for further research. Finally, in the last section, a summary and conclusion are provided.

(5)

1. Literature review

A vast amount of literature has been written on the subject of new firm birth and its relationship to unemployment. As mentioned before, no unambiguous answer has been found. A number of studies find that the presence of high unemployment promotes firm births, whereas other studies find that times of low unemployment coincide with times of many firm births. It appears to be the case that time series analyses point in the first direction and cross sectional analyses show the second result (Storey, 1991).

Although some of the studies cover the opposite relationship than this thesis does and some of them cover the same relationship, valuable insights and techniques can be gained from both directions of research. This section explains these theoretical insights and techniques and explains the base model and its variables.

1.1. Theoretical insights

Van Stel and Storey (2004) provide three reasons from a macroeconomic perspective why high levels of entrepreneurship – and consequently high firm birthrates – could lead to decreases in

unemployment. The first, most obvious, reason they discuss is that new firms add to the stock of jobs, so more people can become employed. This is the main idea of the unemployment pull or

entrepreneurial effect mentioned before. The second reason entails a side effect that comes with the birth of new firms. Van Stel and Storey argue that new firms pose a (real or imaginative) threat to incumbent firms and encourage the latter to perform better. Better overall performance of firms increases economic growth and this in turn decreases unemployment. The final reason van Stel and Storey give is that new firms provide a vehicle for the introduction of new ideas and innovation to an economy. New ideas and innovation have been shown to be an important source of economic growth (Romer, 1986).

It is also possible to look at the relationship between startup firms and unemployment from a microeconomic perspective. Koellinger and Thurik (2012) apply prospect theory – the notion that individuals who are in a gain position relative to their individual reference points are risk averse, whereas individuals in a loss position become risk seeking – to the decision to start a business. They suggest that business ideas that entail risk and uncertainty are more likely to be pursued by individuals who suddenly have lower opportunity costs to self-employment. Examples are people who have become unemployed or people whose salaries have been reduced. Unemployment ‘pushes’ people into self-employment. This effect has been called the unemployment push, refugee or desperation effect.

The microeconomic view illustrates how the relationship between start-ups and unemployment works two ways. On the one hand, unemployment reduces the opportunity costs for individuals of starting

(6)

their own business increasing the number of startups, on the other hand these unemployed people become self-employed and reduce unemployment again..

In a different paper Thurik, Carree, Van Stel, and Audretsch (2008) give an important

counterargument to the unemployment push or refugee effect. In contrast to what prospect theory suggests – that increasing unemployment leads to increasing start-up activity because the opportunity cost of starting a firm has decreased for the unemployed – they say that the unemployed are not likely to start a firm. The main reason for this is that the unemployed tend to possess lower endowments of human capital and entrepreneurial talent needed to start and sustain a new firm. Apart from low levels of human capital, Thurik et al. also mention that unemployment implies lower levels of personal wealth, reducing the likelihood of choosing self-employment. Lastly, high unemployment rates may coincide with low economic growth leading to fewer entrepreneurial opportunities.

1.2. Techniques

As it pertains to technical difficulties that should be taken into consideration, Audretsch & Acs (1994) point out the fact that the most commonly used measure of firm birth in empirical studies has been the change in the number of firms over a given period. This is also called ‘net’ entry. However, by only measuring the change in the number of firms the number of exiting firms remains unrecognized. It could well be the case that the number of new firms is very high, but because the number of exiting firms is even higher, the net entry of new firms is negative. It is therefore important to look at the gross number of newly entering firms and exiting firms and their effects on unemployment separately. Another important note to make is that the effect of start-ups on unemployment may be lagged. In their study on the effects of firm births and deaths in UK counties Johnson & Parker (2006) find a significant (original emphasis) negative effect of births on unemployment. The result is strongly reinforced if a one or two year lag is introduced. At the same time Johnson & Parker also find a positive effect of firm deaths on unemployment, lagged one year. The lesson to take from these findings is that incorporating time lags can increase the accuracy of a time series regression. As starting a business – or discontinuing it – takes time, the effects on unemployment will also take time to manifest. It is not relevant to take into consideration time lags in a cross-sectional analysis.

However, instead of taking data from 1 year, this paper will take average values over 5 years to give a more comprehensive view.

1.3. Baseline model: labor market institutions

To see how start-ups influence unemployment an existing model on unemployment will be extended. Nickell, Nunziata and Ochel (2005) use changes in labor market institutions to explain changes in long-run unemployment. The main reason they offer for choosing labor market institutions to explain unemployment is that the equilibrium level of unemployment is changed first, by any factor that

(7)

influences how easily unemployed individuals can be matched to available job vacancies, and second, by any factor which tends to change wages in a direct manner despite excess supply in the labor market. These factors often take the form of labor market institutions. The labor market institutions used in the paper are the unemployment benefit system, systems of wage determination, employment protection, labor taxes and barriers to labor mobility.

1.3.1. Unemployment benefit system

Nickell et al. mention four aspects through which the unemployment benefit system may influence unemployment, although not all of these aspects are equally measurable. The four aspects are the level of unemployment benefits, the duration of being entitled to benefits, the coverage of the system and the strictness with which the rules of the system are enforced. Of these only the first two can be included in the time series model, because the coverage of the system and strictness of operation are hard to measure numerically. The first aspect that is included in the model, the level of unemployment benefits has a positive effect on unemployment. It determines how necessary it is for the unemployed to find a new. If the level of benefits is high, it is likely that the unemployed individual is able to cover costs of living and the need to find additional sources of income is low.

The second aspect included, the duration of benefits, also has a positive effect on unemployment. The longer the unemployed individual is entitled to unemployment benefits, the longer the need for other income sources is suppressed.

1.3.2. Systems of wage determination

In many European countries the majority of employees have their wages set by collective bargaining between employers and trade unions at the plant, firm, industry or aggregate level. This does not necessarily mean that there are many union members, as often by law, the agreements set by collective bargaining are extended to cover non-members in the same sector. Evidence exists that trade unions and their power to influence wages has an impact on unemployment.

As trade unions tend to set wages above an efficient market level, employers will want to hire fewer employees and more individuals will be unemployed.

1.3.3. Employment protection

Labor market inflexibility is attributed by many to employment protection legislation. Despite this, evidence on the short-term impact of employment protection on unemployment is mixed. There is, however, a clear positive relationship between employment protection and long-term unemployment. An explanation for this could be that employment protection slows down the matching process between the unemployed and positions that are being filled by employees under protection.

(8)

1.3.4. Labor taxes

The influence of labor taxes on unemployment remains subject of debate as some studies show hardly any impact of employment taxes on unemployment, whereas other studies conclude that they can explain virtually the entire rise in unemployment in the period between the 1960s and1990s.

1.3.5. Barriers to labor mobility

Oswald (1997) proposes that barriers to geographical mobility play a key role in determining

unemployment. The ease with which workers can move from one place to another influences the time it takes to match the unemployed with job vacancies. According to Oswald home ownership, relative to private renting, is a key barrier to geographical mobility. Nickell et al. include the percentage of owner occupied housing in their model, but an insignificant effect. The data they use were obtained from the 1997 Oswald study itself and were highly interpolated. Because of the insignificant minimal effect this paper will not include owner occupied housing in the model.

According to Nickell et al. all models on unemployment fundamentally have the same broad implications. One of these implications, which was mentioned above, is that unemployment is determined by how easily unemployed individuals can be matched to available job vacancies and by factors that change wages despite excess supply in the labor market. The labor market institutions cover this implication, but two other broad implications that all unemployment models have in common remain uncovered by the model’s variables thus far. First, short- and long-run unemployment must be consistent with the real demand level and second, real demand and unemployment move towards a level consistent with stable inflation. Since the model is meant to explain actual unemployment, it should accommodate for these and other factors that might cause unemployment to deviate from its equilibrium level temporarily. Therefore ‘shocks’ are included in the model. The variables that are included in this paper to account for these shocks are gross domestic product (GDP) and the real (ex-post) interest rate.

2. Econometric analysis

In this section the methodology of investigating the effect of startups on unemployment is explained and the data used are discussed. This section also contains hypotheses and results of the analysis.

2.1. Methodology

The aim of this paper is to answer the question whether the number of startups in different European countries influences unemployment. The approach is to first see how much of unemployment can be explained by the base model of Nickell et al. and next to see if the fit of the line improves if firm births and deaths are taken into account.

(9)

The model used is a linear regression model (OLS).

𝑼𝑼𝒊𝒊𝒊𝒊= 𝜷𝜷𝟎𝟎+ 𝜷𝜷𝟐𝟐𝑩𝑩𝑩𝑩𝑩𝑩𝑩𝑩𝑩𝑩 + 𝜷𝜷𝟑𝟑𝑩𝑩𝑩𝑩𝑩𝑩𝑩𝑩𝑼𝑼𝑩𝑩 + 𝜷𝜷𝟒𝟒𝑼𝑼𝑩𝑩𝑼𝑼𝑼𝑼𝑩𝑩 + 𝜷𝜷𝟓𝟓𝑷𝑷𝑩𝑩𝑼𝑼𝑷𝑷𝑩𝑩𝑷𝑷 + 𝜷𝜷𝟔𝟔𝑷𝑷𝑻𝑻𝑻𝑻 + 𝜷𝜷𝟕𝟕𝑺𝑺𝑷𝑷𝑻𝑻𝑩𝑩𝑷𝑷𝑼𝑼𝑷𝑷 + 𝜷𝜷𝟖𝟖𝑩𝑩𝑩𝑩𝑷𝑷𝑩𝑩 + 𝜷𝜷𝟗𝟗𝑮𝑮𝑩𝑩𝑷𝑷 + 𝜷𝜷𝟏𝟏𝟎𝟎𝒊𝒊 + 𝒆𝒆

BENRR: Benefit replacement ratio BENDUR: Benefit duration

UNION: Percentage of employees who are union members PROTEC: Degree of employment protection

TAX: Implicit tax rate STARTUP: Startup rate

ENTD: Enterprise death rate GDP: Gross Domestic Product i: Ex post real interest rate

The model by Nickell et al. has been simplified in a few ways. First of all the so called coordination variable is excluded. Coordination refers to mechanisms whereby the aggregate employment

implications of wage determination are taken into account when wage bargains are struck. Nickell et al. use data on coordination from a 2000 paper by Ochel in which existing data is highly interpolated. No recent database exists, so coordination has been cut from the model.

Another change is the elimination of the labor mobility variable. Labor mobility can be reflected by the rate of owner occupied housing. This variable did not yield a significant result in explaining unemployment in the original model.

A third adaptation is the way in which shocks enter the model. The original model includes money supply shocks, shocks to total factor productivity, labor demand shocks, real import price shocks and the ex post real interest rate. The majority of countries included in this study are member of the European Monetary Union (EMU) so money supply shocks and import price shocks are the same across these countries. Total factor productivity and labor demand shocks are captured in changes in GDP. Therefore this model includes GDP rather than the original shocks. The real (ex post) interest rate is still included.

(10)

2.2. Data

The 24 European countries in the sample were chosen based on availability of data on the number of firm births. The countries are:

Austria France Lithuania Romania

Belgium Germany Luxembourg Slovakia

Bulgaria Hungary Netherlands Slovenia

Czech Republic Ireland Norway Spain

Finland Italy Poland Sweden

Denmark Latvia Portugal United Kingdom

These countries have available data on the number of firm births and deaths from 2008 until 2012. 2.2.1. Unemployment benefits

Unemployment benefits enter the regression as the benefit replacement ratio (BENRR) and as an index of the benefit duration (BENDUR). The benefit replacement ratio is the percentage of initial income that is covered by the unemployment benefits. The OECD collects data on the unemployment benefit replacement ratio for three different family types (single, with dependent spouse, with spouse at work) in three different duration categories (1st, 2nd, 3rd, 4th and 5th year). The benefit replacement ratio is equal to the average over different family types in the 1st year duration category. The index of benefit duration is calculated as follows: [0.6(2nd and 3rd year replacement ratio) + 0.4(4th and 5th year replacement ratio)] - (1st year replacement ratio). In other words: the measure of benefit duration is the level of benefit in later years of unemployment normalized on the benefits in the first year of unemployment.

2.2.2. Systems of wage determination

No complete dataset on collective bargaining coverage (the proportion of employees whose wages are set by collective agreements) exists. The variable used as approximation of the influence of collective wage setting is the percentage of employees who are union members (UNION). However, this is only part of the story. Any results should be treated with caution, because trade union density does not reflect the bargaining power of unions. Trade union density rates should always be interpreted within the particular political and social context and according to the legal and institutional framework (Hayter & Stoevska, 2008).

2.2.3. Employment protection

Employment protection is included in the regression through the OECD employment protection index. It captures the strictness of employment protection regulation on a scale from 0 (low) to 6 (high). The measure used is the so-called EPRC_V3 measure. It takes into account 13 indicators on employment

(11)

protection concerning individual dismissals and for collective dismissals. The indicators that are included cover three different aspects of employment protection regulations as in force on the 1st of January each year.

2.2.4. Labor taxes

To include labor taxes into the model data on the implicit labor tax rate from the European

Commission is used. The implicit labor tax rate is obtained by dividing total revenue from labor taxes by the economy wide wage bill.

2.2.5. Startup rate

The startup rate is defined as the number of new firms in a certain time period divided by the average total firm population in that period. For most European countries Eurostat provides data on firm formation from 2008 onwards. Therefore the startup rate is calculated by dividing the average number of startups from 2008 until 2012 by the average firm population from 2008 until 2012.

2.2.6. Enterprise death rate

Similar to the startup rate, the enterprise death rate is the number of discontinued firms in a certain time period divided by the average total firm population in that period. Eurostat data from 2008 until 2012 is used to compute the rate.

2.2.7. GDP

I include in the regression the annual percentage change in GDP per capita. Data is collected from Eurostat for all countries except for the Czech Republic, for which data is obtained from the OECD. 2.2.8. Interest rate

The interest rate used is the Eurostat long-term interest rate on 10-year government bonds.

2.3. Hypotheses

For each variable the direction of the effect on unemployment can be hypothesized. 2.3.1. Benefit replacement ratio and benefit duration

The level of the benefit replacement ratio determines how much of the original salary is replaced by the benefits received. The higher this ratio, the smaller the incentive to find a new job, so the effect on unemployment is expected to be positive. The same holds for the duration of unemployment benefits. The longer unemployment benefits are assigned, the longer unemployed individuals have a source of income and the lower the need for a new job.

2.3.2. Union density

Union density is used as a proxy for collective wage setting. If more wages are set collectively above 11

(12)

the level of an efficient market, firm’s labor costs will rise and fewer people will be unemployed. Therefore the expectation is that union density will have a positive effect on unemployment. 2.3.3. Employment protection legislation

Employment protection legislation is expected to have a negative effect on unemployment. If employees are harder to lay off, fewer people will become unemployed.

2.3.4. Labor tax

Taxes on labor decrease the incentive to work. Therefore it is expected that labor taxes have a positive effect on unemployment.

2.3.5. Startup rate

As was discussed in the literature review, this paper focusses on the unemployment pull effect; that high startup rates coincide with times of low unemployment. Following that reasoning, the

expectation is that the startup rate has a negative influence on unemployment. 2.3.6. Enterprise death rate

Opposite to the startup rate, the enterprise death rate is expected to increase unemployment. If many firms discontinue business, their employees will become unemployed and contribute to a higher unemployment rate.

2.3.7. GDP

GDP is negatively associated with unemployment. Therefore we expect the effect of GDP in the analysis to be negative.

2.3.8. Interest rate

Classic macroeconomic theory states that higher interest rates lead to lower investment spending. If the public sector decreases investment spending, fewer employees are needed to sustain business activities. Therefore it is expected that interest rates have a positive effect on unemployment.

As figure 1 shows, for most countries the enterprise birth rate and the enterprise death rate are quite similar. We would therefore expect the possible effects of these variables to be of similar magnitude

(13)

Figure 1. Enterprise birth versus enterprise death

2.4. Results

Three regressions have been run. In the first regression we try to estimate which part of

unemployment we can explain by simply considering the labor market institutions and shocks from the model by Nickell et al. In the second regression we exclude employment protection to include three more countries with missing data on this particular variable. Finally, in the third regression, we look at how the model performs when firm births and deaths are included in the model.

Table 1. Explaining unemployment

Dependent variable Ut

Independent variables 1 2 3

Constant 0.066 (0.64) 0.027 (0.28) -0.011 (-0.09)

Benefit replacement ratio 0.005 (0.03) 0.057 (0.63) 0.057 (0.57) Benefit duration 0.016 (0.38) 0.030 (0.77) 0.037 (0.84) Union density -0.004 (-0.08) -0.033 (-0.70) -0.025 (-0.47) Employment protection 0.006 (0.19)

Labor tax -0.158 (-0.92) -0.085 (-0.61) -0.041 (-0.24)

Startup rate 0.147 (0.38)

Enterprise death rate 0.054 (0.12)

GDP 0.006 (0.01) -0.123 (-0.22) -0.220 (-0.34) interest rate 1.224 (2.28) 1.044 (2.22) 0.968 (1.74) R2 _0.443 _0.38 _0.3622 N 21 24 24 Notes: a) T-values in brackets

b) Bulgaria, Lithuania and Romania are excluded from the first regression due to unavailable data on employment protection in these countries.

c) For all countries the average startuprate was calculated over 5 years, from 2008-2012, except for Denmark which has data starting from 2009.

d) TheBreusch-Pagan / Cook-Weisberg test was used to see if the error term was homoscedastic. There is insufficient

evidence to reject the null hypothesis of homoscedasticity, therefore, unlike is the case in the original model by Nickell et al. we do not allow for heteroscedasticity.

13 0 0.05 0.1 0.15 0.2 0.25

Enterprise birth rate Enterprise death rate

(14)

Apart from the interest rate, none of the variables seem to have a significant effect on unemployment.1 Many reasons can be put forward for why this is the case. Chapter 3 will elaborate on this subject.

No hard conclusions can be drawn from any of the results whatsoever, but despite the insignificant results we can still look at the direction of the coefficients to see what the cross section analysis suggests. The benefit replacement ratio and benefit duration seem to have a positive effect on unemployment across all three regressions, whereas union density and labor tax have a negative effect. Employment protection was included only in the first regression and has a positive effect. GDP has a positive effect in the regression where employment protection is included, but a negative effect when it is excluded. The interest rate has a positive and significant effect on unemployment across all three regressions.

We can also see how the model compares to reality by comparing the models predictions with actual unemployment. Figure 2. Shows the actual unemployment rates and the predicted rates of the three different regressions. For all three regressions it can be seen that the model generally understates the unemployment rate in countries that have high levels of actual unemployment and that it understates the level for low unemployment countries.

1

To thoroughly check whether a relationship exists, several extra regressions were run leaving out the least significant variables, such as tax and the enterprise death rate. A regression was also run with the change in variables over 2008-2012 instead of average values over 2008-2012. None of these actions led to significant results.

14

(15)

Figure 2. Actual versus predicted unemployment per regression 0 0.02 0.04 0.06 0.080.1 0.12 0.14 0.16 0.180.2 Sp ai n La tv ia Li thua ni a Sl ov aki a Ire la nd Po rtu ga l Hung ar y Bu lga ria Fr an ce Pol an d Ita ly Sw ed en Fi nl an d Be lgi um U ni te d K ing do m Sl ov en ia G er ma ny Ro m an ia De nma rk Cz ech Re pu bl ic N or w ay Lu xe m bo ur g Ne the rla nds Au st ria

Regression 2

Actual unemployment Predicted unemployment 15 0 0.02 0.04 0.06 0.080.1 0.12 0.14 0.16 0.180.2 Sp ai n La tv ia Sl ov aki a Ire la nd Po rtu ga l Hung ar y Fr an ce Pol an d Ita ly Sw ed en Fi nl an d Be lgi um U ni te d K ing do m Sl ov en ia Ge rma ny De nma rk Cz ech Re pu bl ic N or w ay Lu xe m bo ur g Ne the rla nds Au st ria

Regression 1

Actual unemployment Predicted unemployment 0 0.02 0.04 0.06 0.080.1 0.12 0.14 0.16 0.180.2 Sp ai n La tv ia Li thua ni a Sl ov aki a Ire la nd Po rtu ga l Hung ar y Bu lga ria Fr an ce Pol an d Ita ly Sw ed en Fi nl an d Be lgi um U ni te d K ing do m Sl ov en ia Ge rma ny Ro m an ia De nma rk Cz ech Re pu bl ic N or w ay Lu xe m bo ur g Ne the rla nds Au st ria

Regression 3

Actual unemployment Predicted unemployment

(16)

3. Discussion

It is possible that no relationship exists between the variables included in the model in this paper, but there could also be different explanations for the insignificant coefficients on all but one variable. This section will discuss several of these explanations and other limitations to the performed analysis.

3.1. Sample size

One large constraint on the performed analysis is the limited sample size. Data on enterprise birth and death is available from 2008 for only 26 countries, two of which have absent data on other variables. This leaves a sample of only 24 countries. Nickell et al. have only 20 countries in their sample, but they perform a panel data analysis and take into account data from 35 years. Increasing the number of countries in the sample or expanding the analysis to a panel data analysis instead of a cross section analysis could yield more conclusive results. Nickell et al. find significant coefficients for all the labor market institutions they include except for labor market mobility. This analysis finds insignificant coefficients for the startup rate and the enterprise death rate, but also for all the labor market institutions.

3.2. Changes from original model

As was mentioned in the methodology section, some simplifying changes were made to the original unemployment model by Nickell et al.

3.2.1. Different dataset

The data used for this paper differs from the data used by Nickell et al. There is overlap in the countries included in the sample as many countries used in this paper are OECD countries, but the time period included in the two datasets differ substantially. Nickell et al. use older data and include 35 years instead of 5. Not only will different data yield different results, but data from only 5 years will show any possible relationships much less clearly.

3.2.2. Omitting coordination variable

One of the variables that was omitted is the coordination variable. As was mentioned earlier Nickell et al. obtained data on this variable from Ochel (2000b). The OECD data on coordination was highly interpolated. No exact measure like this on coordination exists for the time period considered in this paper. In the original model the coordination variable did show significant influence on

unemployment, as did the interaction terms with coordination (coordination × union density and coordination × labor tax). Excluding coordination and its interaction terms could have led to distorted results.

(17)

3.2.3. Omitting interaction terms

The presence of a significant interaction term indicates that the effect of one explanatory variable on the dependent variable is different at different values of the other explanatory variable. In the original model by Nickell et al. the above mentioned interaction terms with coordination as well as an

interaction term between the benefit replacement ratio and benefit duration were included and significant coefficients were obtained. This paper did not include interaction terms which could possibly explain the different results.

3.2.4. Changing shock variables

In a model that is designed to explain unemployment it is necessary to include factors that can cause short run deviations from the equilibrium level of unemployment. Nickell et al. include five shock variables; money supply, total factor productivity, labor demand, real import prices and the ex post real interest rate. In order to simplify the model, this paper only included changes in GDP and the ex post real interest rate. Distortions between the results of the two studies can be partially due to the difference in included shock variables. Another point to note is the time period included in this paper. From 2008-2012 in most countries unemployment rates were far above their equilibrium level due to the crisis. GDP and interest rates may not have been sufficient to account for this major shock.

3.3. Violation of OLS assumptions

The model used in this paper is the ordinary least squares (OLS) model. Four conditions have to be met in order for OLS to be a good estimator (unbiased and efficient). The conditions are tested for regression 3.

3.3.1. Linearity

The dependent variable, in this case unemployment, needs to be a linear function of the set of independent variables. Nonlinearity is usually evident in a plot of the residuals versus the predicted (fitted) values. The points should be symmetrically distributed around the horizontal zero line. As can be seen from figure 3. the points are not symmetrically distributed, but are more heavily distributed below the zero line than above. There is one outlier. Interpretation is very difficult in such a small sample. A regression of unemployment on the squared independent variables does not lead to a clear cut answer either. As can be seen in data appendix 2, the coefficients of some variables such as benefit duration, the startup rate and the enterprise death rate, increase slightly in significance, whereas others decrease.

(18)

Figure 3. Residuals versus fitted values for regression 3

If linearity is violated this can lead to wrong determinants and estimates or a (non-linear) relationship that actually exists can remain undetected.

3.3.2. Zero conditional mean

The zero conditional mean assumption states that the expected value of the error term is zero, given any value of the independent variables. There is no way of testing if the zero conditional mean assumption holds, as OLS is constructed in such a way that it yields errors with an expected value of zero. This becomes visible when the residuals are plotted against the independent variables. If the residuals have a mean of zero they should be randomly and symmetrically distributed around zero under all conditions. Data appendix 3 shows the plots of the residuals versus all the variables in regression 3. As the plots show, the residuals are mainly distributed along the zero line for the different variables. It is not possible to tell if this assumption is violated, but if it is, this could be due to a violation of the linearity assumption or due to omitted variable bias and it will cause the intercept to be biased.

3.3.3. Homoscedasticity

The homoscedasticity assumption requires the conditional variance of the error to be constant for all variables. If errors are homoscedastic the models uncertainty is equal across all observations. The Breusch-Pagan / Cook-Weisberg test for heteroscedasticity was used. The null hypothesis is that the

(19)

errors are homoscedastic. The P-value of the test was 0.6627, which means that we cannot reject the null-hypothesis.

3.3.4. Normality of the error distribution

The error terms should be normally distributed for the OLS estimator to obtain a t-distributed statistic to test the null hypothesis. Violations of normality create problems for determining whether model coefficients are significantly different from zero. A few large outliers can cause a skewed error distribution. To test for normality of the error distribution a normal quantile plot can be made of the residuals. This plot is sensitive to deviations from normality in the tails.

Figure 4. Normal quantile plot of residuals from regression 3

The plot shows large deviations from normality in the tails. The errors are not normally distributed, which is to be expected in such a small sample. The non-normality of the error term can be clearly seen when a kernel density estimate is made.

Figure 5. Kernel density estimate

(20)

A positive skew can be seen and the kernel density estimate clearly deviates from the skewed normal distribution. The normality of the error term distribution assumption is violated.

3.4. Suggestions for further research

The performed analysis does not show any relationship between the number of startups and unemployment. It does not show any relationship between the labor market institutions from the analysis of Nickell et al. either. A follow up to this analysis could be done by considering panel data on the labor market institutions and the startup rate and enterprise death rate. Data on startups is currently available for no more than five consecutive years, so it is doubtful whether a period this short would yield any results. If data on new firms becomes available for longer periods of time the analysis could yield more insights. Another possible follow up would be to apply a model that allows for reverse causality of unemployment and the number of new firms.

4. Summary and conclusion

This paper has undertaken an analysis of unemployment in 24 European countries from 2008 up to and including 2012. It has attempted to explain unemployment through differences in labor market institutions and differences in the number of startup firms. The aim has been to see if higher levels of new firm formation lead to lower levels of unemployment. To answer this question an existing model on the influence of labor market institutions on unemployment was expanded.

The results indicate that no hard conclusions can be drawn from the performed analysis. The benefit replacement ratio and benefit duration seem to have a positive effect on unemployment, whereas union density and labor tax have a negative effect. Employment protection was included only in the first regression and has a positive effect. GDP has a positive effect in the regression where

employment protection is included, but a negative effect when it is excluded. The startup rate and enterprise death rate both showed a positive effect on unemployment. None of these effects was significant. The interest rate does have significant positive effect on unemployment.

The positive effect of the benefit replacement ratio, benefit duration and employment protection were according to expectation, as was the negative effect of labor tax. Union density was expected to have a positive effect on unemployment, but it showed a negative effect. It was expected that higher GDP would lead to lower unemployment and this showed in the results in two of the three regressions. The startup rate had a positive coefficient, as was expected, but the enterprise death rate also had a positive coefficient, contrary to expectations. The effect of the interest rate was positive as was predicted. The question whether higher numbers of new firms decrease unemployment cannot be unambiguously answered based on the performed analysis. Possible explanations for the insignificant coefficients are the small sample size, various changes to the base model and the violation of OLS assumptions.

(21)

Data appendix

Appendix 1

Country Ut BENRR BENDUR PROTEC UNION TAX STARTUP ENTD GDP i

Austria 0.047533 0.656944 0.739789 2.442177 0.282918 0.408786 0.062404 0.062398 0.0028 0.03444 Belgium 0.07615 0.654111 0.831663 3.001701 0.546405 0.425726 0.05096 0.030519 -0.0034 0.03802 Bulgaria 0.092667 0.751278 0.22007 0.211 0.2507 0.141155 0.105202 0.0168 0.05694 Czech Republic 0.064033 0.684444 0.219156 2.732993 0.154877 0.387496 0.092121 0.089364 0.012243 0.03968 Denmark 0.064067 0.709278 0.499663 2.293084 0.154877 0.348672 0.109237 0.116729 -0.0134 0.02986 Finland 0.076783 0.675889 0.582344 2.166667 0.691259 0.399795 0.095557 0.086425 -0.011 0.03188 France 0.089567 0.713556 0.582716 2.832086 0.076905 0.388089 0.11333 0.068236 -0.004 0.03372 Germany 0.066533 0.713 0.464142 2.977891 0.185069 0.377825 0.085778 0.085151 0.0088 0.0281 Hungary 0.102117 0.642389 0.217937 2.264739 0.12325 0.397647 0.095986 0.124203 -0.007 0.08034 Ireland 0.123633 0.685944 0.750401 1.996599 0.323049 0.264538 0.063146 0.09125 -0.0226 0.06254 Italy 0.08375 0.678722 0.221004 3.032313 0.351098 0.426622 0.069137 0.070994 -0.0188 0.04788 Latvia 0.1521 0.810222 0.176906 2.90703 0.148 0.313856 0.166922 0.114645 -0.0068 0.07922 Lithuania 0.132583 0.599556 0.227947 0.093333 0.320332 0.210823 0.174041 0.0148 0.07034 Luxembourg 0.049 0.852333 0.174423 2.735261 0.348438 0.320866 0.095848 0.07525 -0.0228 0.0335 Netherlands 0.047667 0.725833 0.452721 2.893424 0.184732 0.371442 0.10548 0.070966 -0.0058 0.03166 Norway 0.0619 0.696944 0.492627 2.309524 0.533446 0.360804 0.085704 0.06089 -0.0068 0.03444 Poland 0.089317 0.5705 0.245399 2.39059 0.141674 0.317595 0.127451 0.106232 0.032 0.05786 Portugal 0.120183 0.774222 0.6688 3.307823 0.200091 0.244517 0.127295 0.168067 -0.0102 0.06984 Romania 0.066067 0.510944 0.267479 0.328 0.298878 0.117826 0.116817 0.0132 0.0774 Slovakia 0.1277 0.694944 0.209129 2.54093 0.169758 0.320471 0.139206 0.107278 0.0198 0.0446 Slovenia 0.069383 0.75375 0.192814 2.68254 0.254059 0.353742 0.10699 0.078536 -0.0138 0.0472 Spain 0.190467 0.709278 0.455017 2.619048 0.174891 0.325564 0.077223 0.094765 -0.015 0.04778 Sweden 0.0776 0.621222 0.613414 2.517007 0.679892 0.394374 0.073555 0.061532 0.0018 0.02846 United Kingdom 0.073833 0.4715 0.534464 1.713152 0.263969 0.255063 0.113967 0.116485 -0.0142 0.03166

Appendix 1. Data summary

(22)

Appendix 2

Dependent variable Ut

Independent variables 1 2 3

Constant 0.102 (1.59) 0.061 (1.11) 0.034 (0.50)

Benefit replacement ratio2

-0.025 (-0.26) 0.043 (0.67) 0.05 (0.73) Benefit duration2 0.001 (0.01) 0.024 (0.57) 0.038 (0.78) Union density2 0.017 (0.25) -0.015 (0..25) -0.01 (-0.16) Employment protection2 _0.002 _(0.42) Labor tax 2 _{-0.333 (-1.19)} _{-0.164 (-0.7)} _-0.083 _(-0.30) Startup rate2 0.8 (0.58)

Enterprise death rate2

0.2 (0.10) GDP2 _{-6.140 (-0.16)} 7.565 (0.21) 14.485 (0.38) interest rate2 _10.650 _(2.17) _{8.834 (2.01)} _7.58 _(1.45) R2 _0.421 _0.327 _0.355 N 21 24 24

Note: T-value in brackets

Appendix 2. Unemployment regressed on squared variables

(23)

Appendix 3

Appendix 2. Residuals plotted against independent variables of regression 3

(24)

References

Literature

Audretsch, D. B., & Acs, Z. J. (1994). New-firm startups, technology, and macroeconomic fluctuations. Small Business Economics, 6(6), 439-449.

Evans, D. S., & Leighton, L. S. (1989). Some empirical aspects of entrepreneurship. The American

Economic Review, 79(3), 519-535.

International Labour Organization. (2009). Social Dialogue Indicators: International Statistical Inquiry 2008-09 (technical brief). Geneva, Switzerland: Hayter S. & Stoevska V.

Johnson, P., & Parker, S. (1996). Spatial variations in the determinants and effects of firm births and deaths. Regional Studies, 30(7), 679-688.

Koellinger, P. D., & Roy Thurik, A. (2012). Entrepreneurship and the business cycle. Review of

Economics and Statistics, 94(4), 1143-1156.

Nickell, S., Nunziata, L., & Ochel, W. (2005). Unemployment in the oecd since the 1960s. what do we know? The Economic Journal, 115(500), 1-27.

Ochel, W. (2000). Collective bargaining (centralization and co-ordination). Ifo Institute, Munich. Romer, P. M. (1986). Increasing returns and long-run growth. The Journal of Political Economy,

94(5), 1002-1037.

Storey, D.J., 1991, The Birth of new firms – does unemployment matter? A review of the evidence. Small Business Economics, 3(4), 167-178.

Thurik, A. R., Carree, M. A., Van Stel, A., & Audretsch, D. B. (2008). Does self-employment reduce unemployment?. Journal of Business Venturing, 23(6), 673-686.

Van Stel, A., & Storey, D. (2004). The link between firm births and job creation: is there a Upas tree effect?. Regional Studies, 38(8), 893-909.

Data

European Commission (2015). Tax and benefits database. (Accessed on 19 June 2015).

Eurostat (2015). Table [bd_9bd_sz_cl_r2] – Business demography by size class (from 2004 onwards, NACE Rev. 2). (Accessed on 18 June 2015).

Eurostat (2015). Table [gov_tax_a_itr] – Implicit rates by economic function. (Accessed on 19 June 2015).

Eurostat (2015). Table [nama_aux_gph] – GDP per catpita – annual data. (Accessed on 18 June 2015).

Eurostat (2015). Table [une_rt_a] - Unemployment rate by sex and age groups – annual data. (Accessed on 18 June 2015).

ICTWSS (2012). Trade union density. (Accessed on 22 June 2015).

OECD (2015). Gross domestic product (GDP) (indicator). (Accessed on 18 June 2015).

(25)

OECD (2015). OECD Indicators of Employment Protection. (Accessed on 19 June 2015). OECD (2015). OECD Trade union density. (Accessed on 19 June 2015).

OECD (2015). Social policy indicator database. (Accessed 19 June 2015).

The effect of start-up firms on unemployment : in 24 European countries from 2008-2012