Does entrepreneurship pay in Germany?

Victor Hugo Borja Bustamante Student ID 10824464 Thesis Director Dr. Philip Koellinger 1st July 2015

MSc. in Entrepreneurship

-

MASTER THESIS

-

ACKNOWLEDGEMENTS

I wish to express my sincere thanks and admiration to Dr. Philip Koellinger for providing me with all the necessary support and knowledge to complete this thesis. Moreover, I want to take this opportunity to thank the University of Amsterdam for believing in me. I am also grateful with Ramón Alberto, Arne, Patrick, Quentin and Umer for their friendship and valuable advice.

A special thanks to my family. Words cannot express how grateful I am to my mother Patricia, father Victor and sister Patty, for all the sacrifices that they have made on my behalf.

At the end I would like to express appreciation to my beloved girlfriend Sarah, who spent sleepless nights with and has been always my unconditional support. I would not have made it this far without her.

- 3 - Table of contents Abstract ... - 4 - 1. Introduction ... - 5 - 2. Literature Review ... - 7 - 3. Theoretical models ... - 10 -

4. Challenges and limitations in the literature ... - 11 -

5. Data description ... - 12 -

5.1 The German Socio-Economic Panel ... - 12 -

5.2 Data Collection ... - 13 - 6. Data Analysis ... - 15 - 6.1 Control Variables ... - 15 - 6.2 Measuring Earnings ... - 17 - 7. Methods ... - 19 - 7.1 Panel Data ... - 19 - 7.2 Pooled OLS ... - 20 - 7.3 Quantile Regression ... - 21 - 7.4 Fixed-effects ... - 21 - 8. Results ... - 23 - 8.1 Descriptive statistic... - 23 - 8.2 Empirical Results ... - 24 - 9. Discussion ... - 32 -

9.1 Implications for theoretical models ... - 32 -

9.2 Limitations ... - 33 -

10. General Conclusions ... - 36 -

References ... - 37 -

- 4 -

Abstract

Financial returns are undoubtedly one of the most important reasons why people chose a career. Entrepreneurship phenomenon has recently taken control over the minds of people, media and research. The present study pretends to find an answer to the questionable returns to entrepreneurship. This research follows one of the most important studies taken in the subject “Does

entrepreneurship pay” by Hamilton (2000). Previous studies have found that

entrepreneurs have lower financial returns than paid employees. Several theories have been proposed to explain the earning difference and the compensations to lower returns. Using the German Socio-Economic Panel Study (SOEP) from 2010, panel data is analyzed with several statistic and econometric technics in order to understand and evaluate the earning differentials. Findings suggest that in Germany, individuals have higher earnings by being entrepreneurs.

- 5 - 1. Introduction

Entrepreneurship is a phenomenon that has taken importance among scholars, policy makers and society during the last decades. As it can be studied from different perspectives, multidisciplinary research teams have provided over years a set of definitions, models and explanations around it. Despite the academic effort, entrepreneurship has still an infinite number of puzzling questions due to the dynamics and complexity of its nature. Therefore, it is important to adapt each of the theoretical proposes to the context of the research to avoid jumping into generalized conclusions.

Expected monetary compensation and social values are some of the most important detonators for choosing a career. According to the Global Entrepreneurship Monitor in 2014, European economies showed the lowest social values towards entrepreneurship in all dimensions: career choice preference, status in society and media attention. For instance, Germany has a low percentage of people who want to become entrepreneurs, compared to other innovation driven countries such as the U.S. Despite that, it seems that being an entrepreneur in Germany is not an attractive career choice. However, successful entrepreneurs are highly recognized and respected in their society.

While entrepreneurs can be seen as “superheroes” due to high media attention in countries like the U.S., researchers have moved idolatry aside and provided empirical evidence for the financial returns in entrepreneurship. Entrepreneurship is a high risk activity in which most of its owners tend to invest their personal wealth into their own business. On the other hand, wage workers have a higher chance to minimize risk by diversifying their wealth into investment portfolios. Therefore, higher returns on entrepreneurship would be expected to support the risk associated to it. Evidence found that returns are no different from those held by wage worker portfolios (Moskowitz & Vissing-Jorgensen, 2002), however. In fact, many academics (i.e.(Albarran, P., Carrasco, R., Martinez-Granado, 2009; Kawaguchi, 2002; Lazear & Moore, 1984) have concluded that self-employment has a negative financial impact for most of the people who enter in this job sector as returns tend to be less than if they stayed or became paid employees.

Theoretical models have been proposed to explain why individuals would remain or even look for self-employment even though the returns are negative. Non-pecuniary benefits such as “being your own

- 6 -

boss” or “higher satisfaction” are some of the findings attributed to it. One of the most famous studies which examine the differences in the earnings distributions of paid and self-employed individuals was conducted by Hamilton (2000) with a U.S. population sample. The aim of this paper is to replicate in essence the paper “Does entrepreneurship pay?” by Hamilton (2000) and contextualize it exclusively to the German male working population. For this purpose, the German Socio-Economic Panel has been chosen. External validity is not the aim of this research as social, cultural and political differences may exist between databases. The research will be limited to specific aspects of the SOEP 2010 panel. Moreover, the replication of Hamilton´s study (2000) will be complemented with additional research methods such as fixed-effects in order to control for unobservable factors. Even though there are certain frameworks encircling the definition of entrepreneurship, it is still highly associated with self-employment as they share in essence the same characteristics. For the conceptualization of this research, entrepreneurs and self-employed will be taken as synonyms.

The question is open to everyone interested in this research, does entrepreneurship pay in Germany? The remainder of this paper will be structured as follows. First, a review of the most relevant literature in the subject, including Hamilton’s findings, will be presented. Next to it, a description of the theoretical models predicting earning differentials will be given. Second, the German Socio-Economic Panel will be introduced and the logic for obtaining the final database used in this study will be discussed. Third, the relevance of the variables analyzed for this study will be presented. Forth, econometric methods and their research potential are presented. Fifth, empirical data and its analysis will be presented, followed by the discussion and limitations associated to it. Finally, general conclusions will be held.

- 7 -

2. Literature Review

Income distribution among entrepreneurship is a very interesting topic as the variable represents success in western economies and nonetheless, it presumed to be a determinant of occupational choice (Poschke, 2013). During the last years and with the increasing attention to entrepreneurship, scholars have done research around the topic. Most importantly, contradictory results have been found. In the next section, a review of the most relevant evidence will be presented.

One of the most complete and famous research concerning the returns in entrepreneurship was conducted by Barton Hamilton (2000). The author compiles theory and empirical evidence on the financial return differences that exist between entrepreneurs and wage workers. Hamilton (2000) presents an empirical research based on the 1984 U.S. panel of the Survey of Income and Program Participation (SIPP). With the data of this panel, Hamilton (2000) constructed wage worker/self-employed profiles to measure income differentials. The panel was built with nine quarterly waves from 1983 to 1986. There were 8,771 male school leavers aged from 18-65 working in the non-farm sector. The objective of the research was to determine “the extent to which the behavior of workers choosing

to enter or remain in self-employment could be explained first by Investment and agency models; second, matching and learning models and finally; working conditions like non pecuniary benefits”(Hamilton, 2000).

One of the main characteristics of Hamilton’s (2000) research is that he measured income in three different ways: net profits, draw and equity adjusted draw (EAD). He presents these other two alternative measurements to compensate for those limitations associated to net income measurements. A detailed explanation of these limitations will be given further in this paper. In a few words, “draw” is the amount of consumption the business generates for its owner and “EAD” is the sum of the draw in period t and the change in business equity between the beginning of period t and period t+1. EAD takes into consideration the returns of entrepreneurship in firm value (returns to capital). Some businesses are started with the objective of being sold in the future to harvest the investments, human and monetary. In this sense it is important to measure it as it may be an important reason why people enter entrepreneurship. The alternative measures are expected to have less underreporting problems as

- 8 -

their components are not reported to tax authorities, increasing the value of the results. However, the different measurements were able to set due to the fact that the SIPP reports key information to build them.

Ordinary least squares (OLS) and quantile regressions were conducted as part of the analysis due to the skewness observed in the data. Moreover, the conclusions that can be drawn from analyzing the earning differentials vary with the measures that were used. A cross-sectional analysis shows that the mean earnings are lower for self-employed if net profit and draw are taking into consideration. EAD apparently has a higher mean estimation for self-employed, however this is only true for the top 25% of the entrepreneurial population. Based on these results and in the fact that starting a new business is a risky activity due to the high failure rates (Jones & Jayawarna, 2010), some people would conclude that it is better, at least for 75% of the people, to become or stay wage workers (Astebro, 2010) . For net profit and for draw interpretation is indisputable, the earnings are always lower for self-employed compared to wage workers. Going back to the importance of the earning distribution, the tables and graph presented showed that the variance (standard deviation) for all measures of self- employment income were significantly higher than for wage worker, between 2 and 3.5 times.

Given this situation, Hamilton (2000) focused his analysis on the median rather than the mean incomes to avoid distortions arising from the long upper tail income distribution. Nonetheless, the presented conclusion did not change when using multivariate regressions. Hamilton (2000) constructed wage profiles in order to observe the impact of labor market experience and job or business tenure on earnings. According to that, paid employment has greater initial earnings and growth potential than self-employment. For instance, by analyzing the entrepreneurs after 10 years in business, median earnings were 35 percent lower that on paid jobs for the same duration. Entrepreneurial earnings are always less than the predicted starting wage (for zero job tenure) available from an employer, regardless of the length of time in business (Hamilton, 2000). The only case in which entrepreneurs took over paid employees was in the 75th percentile where it is clear that there are substantial returns. Moreover, Kawaguchi (2002) replicated Hamilton’s work based on the US National Longitudinal Survey of Youth 79 (NLSY79) and a sample from 1985 to 1998. He restricted his study to white males as it is commonly done in other studies (i.e. Hamilton, 2000). He measured the differences in earnings between self-employed and wage workers at several points during job market experience and job tenure. According to OLS regression, self-employed earning-experience/tenure profiles are flatter and

- 9 -

as a consequence individuals in this sector tend to earn 18% less than their salaried counterparts, depending on their educational background and marital status. On the other hand, when controlling for individual unobserved characteristics using a fixed effects regression, earnings among job sectors appear to be almost the same.

Even though a substantial body of research concluded that entrepreneurship does not pay (e.g. Borjas and Bronars 1989; Evans and Leighton 1989; and Hamilton 2000) more recent literature has opponent findings. In 2009, academics from the University of Amsterdam and the international Institute for the study of Labor (IZA) looked on the effect of human capital on the performance of entrepreneurs vs wage workers. The dependent variable income was controlled for demographics, education level and general ability. They found evidence that the mean, median and standard deviation of income for self-employed individuals were higher than those for wage workers (Praag, Witteloostuijn, & Sluis, 2009). These results are in line with those of Parker (2004) and Fairlie (2005). Moreover, these results are not explained by professional entrepreneurial labeling such as being a lawyer or a doctor. They concluded that the occupational choice behavior was inconsistent with the higher return to education for entrepreneurs, assuming that labor market decisions are based on income maximization.

Furthermore, Levine & Rubinstein (2012) researched on the returns of entrepreneurship making a clear difference between incorporated and unincorporated wage workers and entrepreneurs. The results show that entrepreneurs earn much more per hour and work much more hours than their salaried and unincorporated equivalents (Levine & Rubinstein, 2012). Differences in the income distribution average 48% more in favor for those incorporated self-employed than wage workers and unincorporated. Interestingly, and in some point in contradiction to Hamilton (2000), Levine et al (2012) manifest that a person who chooses to become incorporated self-employed earns on average 18% more than he was earning as a salaried employee and 6% if the median is compared, concluding in that sense that entrepreneurship does pay.

As it can been observed, there are different positions regarding the monetary returns to entrepreneurship. However, evidence arrives from different contexts and methodologies, though it is no possible to give a final conclusion in a global context. But why would entrepreneurs be willing to get lower returns for their work? Evidence found by a research lead by Taylor (1996) show that higher expected earnings in self-employment relative to wages are key incentives for entrepreneurship. Not the least, he found that a bigger proportion of entrepreneurs give less importance to payment and

- 10 -

security aspects when choosing their job compared to wage workers. That said, it is important to remark that money is not the only or most important incentive for entrepreneurs. Non-pecuniary motivations are not new and can be traced back to the work of Schumpeter (1934) where he recalls the need of entrepreneurs for “success” or to “prove themselves superior to others” (Parker, 2009). Moreover, non-pecuniary benefits such as independence, “being your own boss” (Hamilton, 2000)or job satisfaction (Blanchflower & Oswald, 1990; Kawaguchi, 2002) are some of the more frequent reasons why people decide to switch from paid employments to entrepreneurship. Kawaguchi (2002) analyzed job satisfaction scores confirming that self-employed individuals are more satisfied with their jobs than wage workers. Moreover, one dollar of earnings for self-employed is equivalent to 2.5 dollars of earnings for wage workers, being a 250% difference in terms of job satisfaction. The author concludes that difference is big enough to compensate for and explain the lower earnings among self-employment.

3. Theoretical models

The explanations and interpretation for most of the income differences between self-employed and wage workers fall into different rational choice models. Investment, agency, matching, learning models and working conditions will be described in the following section. Nevertheless, cross-sectional earning differentials may exist across sectors as there are differences in the tenure-earning profiles between sectors (Astebro, 2010).

Investment models argue that companies pay to their employees not for general human but for

firm-specific investments (Becker, 1964). After some years, employees’ turns to be locked into the companies as switching cost are significant. On the other side, with the increase of tenure, employers can reduce the wage of the employee under his marginal product. This phenomenon does not occur with self-employment as the entrepreneur himself is in charge of paying for all human capital investments. Based on this model, wage worker income would have a flatter tenure-earning profile opposed steeper one of the self-employed. Similar to the investment model, agency models (Lazear & Moore, 1984) are based on the belief that income differences arise across sectors. In this case, firms initially pay to their employees less than their marginal product. Salary will increase based on the tenure of the employee. In this way employers will be motivated to stay at the company and this last

- 11 -

one will save money from those employees that are not committed to stay. Larzer (1981) made a comparison between the growth of the earning profiles across sectors and found a steeper income behavior on wage workers than for self-employed. Evidence from agency models contrasts investment model ones as the tendency on the income-profiles is opposite. Matching and learning models (Jovanovic, 1979) highlight that earning differential rise from differentiating working sectors between self-employment and wage workers and their heterogeneous specific sector abilities. Last but not least, earning differentials may be a result of working conditions across sectors. As previously discussed, Hamilton (2000) argues that self-employed workers benefit from non-monetary aspects of entrepreneurship, however his claim is not sustained by empirical evidence as later by Kawaguchi (2002).

4. Challenges and limitations in the literature

“Nothing is perfect. Life is messy. Relationships are complex. Outcomes are uncertain. People are irrational” (Hugh Mackay, n.d); and so is research without the necessarily cautions. Most of the reviewed literature in the last section presented some limitations that have to be taken into consideration before making any critical judgment or conclusion about the results.

Parker (2004) documented some of the reasons that limit the income comparability of the entrepreneurs and wage workers. Not surprisingly, one of the most notable reasons is in fact the income measure itself. Entrepreneurs have more opportunities to underreport their taxable income than employees as the second ones are not in charge of their salary, but the company that they work for. In contrast, self-employed people tend to evade the income tax as they are responsible for reporting the income themselves to the tax authorities (Parker, 2004). Hamilton (2000) managed to construct alternative measures, EAD and Draw, to reduce the uncertainty about the income reported with net profit. Moreover, the EAD measure covers the returns to capital, a characteristic that is only associated to entrepreneurship. However, and as discussed before, this is only possible to construct if the dataset provides with enough information. Other potential limitation in the net income measure is that it does not take into consideration “fringe benefits” which in most of the cases are not taxable. These benefits are only available to employees and sometimes represent a major contribution to their welfare. As examples of these benefits, there are health insurance, matching retirement contributions and life

- 12 -

insurance. For these reasons, the comparison of the absolute income levels should be made with caution (Praag et al., 2009). On the other hand, Superstar theory (Rosen, 1981) suggest that in certain kinds of economic activities, a small number of people earn significant higher amounts of money compared to the mean, particularly for those who are self-employed. Plotting a distribution graph clarifies the distinction of superstars from the rest as the very top rewards skew the income distribution. In this sense, it is hard to compare the mean income of self-employed with the ones of wage workers as the salaries of “superstars” are not representative of the entire population of entrepreneurs and do affect its mean. Last but not least, the availability and reliability of long datasets represents a limitation in a way that the timeframe should produce more accurate results than a cross-sectional analysis. According to Hamilton (2000), the length of time is essential to determine whether there are true income differences between entrepreneurs and wage workers or whether there is only a business-specific distribution across sectors.

5. Data description

5.1 The German Socio-Economic Panel

The main objective of the paper is to analyze and discuss the empirical evidence generated by replicating Hamilton’s (2000) “Does entrepreneurship pay” research. It is not pretend to validate the results, but to compare the evidence by using the same methods, until its possible extent, with a different dataset, however. For this purpose, the German Socio-Economic Panel Study (SOEP) was selected. SOEP is a representative longitudinal study of private households in Germany which has collected data from 1984 to date. As many significant changes have occurred during the last decades (i.e. the reunification of East and West Germany in 1989), a highly compressed and pooled version of SOEP 2010 was released, SOEPLong 2010. This version of the database is characterized by being in “long” format instead of wave-specific individual files. SOEPLong 2010 has harmonized some of the variables that are important for the study, such as income information in euros. All the documentation of the modifications can be found on the official website. The SOEPLong panel consists of the collection of 66,334 individuals and 578,014 observations (interviews). The data panel was chosen as it contains information of individuals over 27 years. This time-frame allow to provide extensive

- 13 -

evidence about the earning differentials among self-employed and wage workers. Additionally, the big amount observations increase the power of the sample, making the statistical data analysis more precise as a major proportion of the population is taken into consideration in the sample

The sample for this study consists of 17,051 males aged from 18 to 65 years. Individuals have been interviewed on a yearly basis from 1984 to 2010. Therefore, the maximum number of observations per individual is 27. The sample consists of 117,483 observations (interviews), on average 6.9 interviews per individual. This implies that most individuals were not interviewed along the entire time-frame of the panel. The final sample came up by applying some filters to the original SOEP dataset. Major steps will be named in the following section for replication proposes if needed in the future.

5.2 Data Collection

As much as the existing literature, individuals who work in farm-related activities were taken out of the sample and their income is considered to be closely related to the governmental subsidies that are provided to them (Hamilton, 2000). Besides the farm-related occupied individuals were present in the original sample, the number was not substantial. A major filter was used to exclude woman from the panel. Women are considered to have disadvantages in the labor market compared to man and are less likely to become entrepreneurs (Parker, 2009). The intention of the filter was to avoid making false inferences about the earning distribution between genders and stick to the method of Hamilton’s. This filter represented a reduction of 50.44% percent of the observations contained in SOEP. Moreover, age range was set from people between 18 -65 years old. The bottom age, 18, corresponds to the official adult age and the top age, 65, to the most common retirement age. This age range also fits the one proposed by Hamilton (2000).

In this paper, opposed to Hamilton (2000), highly paid professional occupations such as doctors and lawyers were kept in the panel as their salary was not top coded. As Praag et al. (2009) concluded in their study, earning differentials cannot be explained by professional entrepreneurial labeling such as being a lawyer or a doctor. In the same line of reasoning “professionals” are taken into consideration as part of the nature of job sectors; enough reasons to keep their data in the analysis. Only one outlier was removed from the sample as the observation was too extreme and could provide non representative information for the group. This observation had a considerable impact on the mean as it

- 14 -

weighted a 2600% of the value of the mean. By considering it on the sample, the year 2010 mean would be inflated by 7%. Therefore, the analysis of this paper will not only focus on small business owners but on a more broad population. To generate the sample, there were only considered those individuals who either had only income from self-employment or a wage. For individuals who were wage workers and self-employed at the same time, as no additional variables could indicate which their main working sector was, for instance, the number or hours worked in each job, were available. Consequently a definition for self- employment and paid job could not be provided and a filtering command had to be applied. This event had no significant weight as these particular individuals represented less than 1% of the original panel. On the other hand, individuals who reported zero income for both, self-employment and paid job, were taken out of the sample as no further analysis could be run over their observations. Additionally, zero individual earning observations were sorted out. Taken this action, observations for unemployed were also deleted. According to Levine & Rubinstein (2012), dividing self-employment into incorporated and unincorporated can describe the earning differentials across sectors. In the particular case of SOEP, a clear distinction was not able to do, reason why it is assumed that both self-employment status are present in final sample. This observation is not limitative as Hamilton included incorporated and unincorporated self-employed individuals in his study.

Two last filters were applied to obtain the final dataset and to provide a definition for the job sectors. First, individuals who had an employment status of part-time were not considered. By doing this, the sample was reduced by 15% but more importantly, “hobby entrepreneurs”. (i.e. Fairlie, 2005; Praag et al., 2009) who are self-employed individuals who work less than 300 hours a year in their businesses were filtered out automatically as low yearly working hours were attributed mostly to part-time workers. Second, people who earned less than 1€ per hour were treated as “casual workers” (close to Hamilton’s definition), a thus, taken out of the sample.

Therefore, self-employees are defined as labor market participants who work more than full time in their own business. Wage workers are defined as persons whose main occupation is a salaried job. Both sectors share same filters in order to have an equally comparable sample. The final dataset consists of 117,483 observations (individual interviews) which correspond approximately to 20% percent of the original panel.

- 15 -

6. Data Analysis

Hamilton’s (2000) research was based on a sample of individuals with similar observable qualifications participating in the SIPP 1984 panel which was covered from 1983 to 1986. Earning differences were obtained by constructing alternative measures for entrepreneurial income to deal with underreporting issues; salary for wage workers; net profit, draw and EAD for self-employed. Moreover, Hamilton (2000) constructed earning profiles based on the amount of years an individual remained in his job (tenure) and his potential work experience. Furthermore, he used longitudinal data to explore the hourly earnings fluctuations associated with switching job sectors. Last but not least, he performed a fringe benefit analysis with health insurance procurement and provided evidence for working-conditions theory. For the current research, a bigger panel database which covers from 1984 to 2010 was selected. The sample was carefully chosen based on most of the observed characteristics Hamilton (2000) proposed for his analysis. Some of these characteristics will be further analyzed in order to comprehend their impact on the earnings distribution of each sector. However, the analysis will be limited to salary and self-employed earning measures as SOEP does not provide the data required to construct alternative ones. Moreover, earning profiles are not constructed. To account for the heterogeneity associated with switching job sectors, longitudinal data was also used in the analysis but with a fixed-effects method which will be explained in the next section. Despite, not all assumptions and considerations were able to follow, valid alternative analysis is presented.

6.1 Control Variables

“Marital Status” is a categorical variable that is composed out of 6 values. For replication purposes, a secondary variable was created, “Recode Marital Status”. This last variable was constructed as a dummy variable were 0 represented “spouse not present” and 1”spouse present”. Only people who had been categorized as “married” in “Marital Status” were identified with the “1” value, all the others were coded as “0”.

“Potential labor market experience” equals the age of the individual minus his years of schooling minus six. A “zero” value was assigned to those observations (108) that had negative computed value. A quadratic expression of this variable is also used as a control regressor.

- 16 -

In order to compare the productivity and demographic characteristics between paid and self-employed individuals, Table 1.A was created. Notice that the data is only representative for the year 2010 as a 27 year period table was not able to be calculated as the sample presents unbalanced data. Nevertheless, it is interesting to see how these characteristics change over time, reason why comparable tables from previous years can be found in the appendix section of this paper. Not surprisingly, for 2010, only 12% percent out of 4,365l observations were attributed to self-employment. Noticeably, self-employed individuals have on average greater potential labor market experience than wage workers do. These first ones are also older, have one more year of education and present a lower rate of disability among the individuals, in average. As for the variable “Marital Status”, it had to be recomputed for replications and interpretation purposes. It can only be inferred that a bigger proportion of entrepreneurs are married and living together with their spouse.

TABLE 1.A

Variable Description, Summary Statistics & Independent Samples T-test by Employment Sector in 2010 Mean T-test Variable Name Description Wage Worker Self-Employed Difference Standard error WORK_EXP Potential labor market experience in

years 26.26 28.78 -2.52*** 0.44

YEDU Number of Years of Education 12.8 13.77 -0.96*** 0.04

AGE Age of the individual 44.83 48.46 -3.63*** 0.43

REC_MS Marital Status, spouse present 0.66 0.68 -0.02** 0.02

RACE Race is non- white - - - -

DIS_STAT Disability status of individual 0.06 0.04 0.02*** 0.011

Observations 3,846 519

* p<0.05, ** p<0.01, *** p<0.001

Before conducting the independent samples t-test to compare the means of each of the regressors for both job sectors, the inferential statistic Levine Test was used to evaluate the equality of variances. Statistical software usually assumes homoscedasticity; however the Levine test provided evidence of the contrary in the SOEP dataset for most of the regressors. Mean differences are not a problem when testing the means without accounting for equality of variance, but the standard errors would be bias. Based on this check, a t-test of equal or unequal variance was performed. Results from the independent samples t-test indicate that there is a highly statistical significant difference between the mean of the

- 17 -

two job sectors. The negative sign in the “difference” is attributed to the coding of the job sector variable were “1” represents employment and “0” paid employment and to the fact that self-employed has higher means. Only disability status had a greater value for wage workers as commented before.

Some limitations were found in the data to match Hamilton’s’ as there was no information available in the SOEP panel. For instance, literature acknowledges evidence that there are negative working factors involved in ethnic minorities (i.e. Borjas & Bronars, 1988; Parker, 2004). These negative factors tend to be related to some kind of discrimination to non-white people which in consequence affects negatively the income distribution. According to (Borjas & Bronars, 1988) minority self-employed workers tend to earn less on average than their white counterparts. For these reasons, researchers suggest to keep the sample and its analysis only within white people. However, it was not the case for this study as SOEP does not record “race” variables as most of the US and UK datasets do.

6.2 Measuring Earnings

In general, empirical research uses net profit as the main measure of performance and earning distribution (Praag et al., 2009). However, income from self-employment and paid work were the only variables available to measure earnings in this study. No other valuable variables were found, reason why alternative measures like Draw and EAD, which try to overcome some of the issues of the income measurement, were not possible to construct. SOPE measures income variables as follows:

Wages, Salaries from main job (IWAGE): Represent an individual unit of observation for salaries from

main job. The measurement is the product of the number of months that the income was received in the previous year and the average gross amount per month. Therefore, it is an annual measurement and it is in euro prices of 2010.

Income from employment (ISELF): Represents an individual unit of observation for

self-employment income. The measurement is the product of the number of months that income was received in the previous year and the average gross amount per month. It has the same construction than wages.

- 18 -

Despite the annual income was available for both wage worker and self-employed, an hourly based variable (HEARNINGS) was constructed in order to avoid conflict with the supposition that self-employed work more than wage workers. In this way, hourly earnings will allow focusing the analysis on the earning differential rather than in hours worked (Hamilton, 2000). Therefore, this paper will focus on explaining hourly earnings differentials.

Figure 1 shows the empirical income distribution on the hourly earnings for each of the job sectors. As it can be observed, a lack of symmetry is present for both distributions and in accordance with Hamilton (2000); the central tendency is less notorious for self-employment. Both job sectors have positively skewed distribution were most of the observations are concentrated in the lower end. Greater dispersion is visible in the distribution of self-employed. The long tails can be attributed to those highly paid wage workers but particularly to those few superstar entrepreneurs who have outstanding hourly earnings. For these last ones, earnings can sometimes reach almost double than the best paid wage worker. This is no reason to make any final conclusion of the results, however.

0 .0 2 .0 4 .0 6 D e n si ty 0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150 Hourly Earnings

Self-employed Wage Worker

- 19 -

7. Methods

7.1 Panel Data

Panel data is more complicated but interesting than simple cross-sectional analysis. Panel data or longitudinal data will provide repeated measurements at different points in time on the same individual units (Adrian Colin Cameron & Trivedi, 2009). By doing this, each observation will provide information of wages, working hours, years of education etc., for the same group of individual across time. In this sense, panel methods are able to capture variations over units, similar to cross-sectional data, but also variation over time. Another advantage of using panel data is that it allows to inquiry into variables that go unnoticed in simple cross-sectional studies as variability is mistreated. Finally, panel models provide bigger samples which allow testing the impact of a larger number of explanatory variables of the level and change in the explained variable (Podestà, 2002).

There are many basic considerations that have to be taken into account when managing panel data. First of all, panel data can be balanced or unbalanced. The distinction between the two of them relies on continuity of the observations. For balanced panels, all individual units are observed in all time periods while in unbalanced, one or more individual observation are missed over time. Moreover, in longitudinal data, standard errors have to be corrected as for every individual each additional year or period of time is not independent of previous periods. In a country level panel analysis, such as in this research, correlations can also be present across individuals. The type of regressors can also influence the regression coefficients for some estimators. For instance, there are time invariant regressors (xit=xi for all t) such as gender or individual invariant (xit= xt of all i) regressors such as a time trend.

Nevertheless, some regressors may also vary over time and individuals. Panel data is highly valuated as it offers the possibility to control for unmeasured individual heterogeneity which would otherwise bias the results.

A very general linear model for longitudinal data can be computed as follows:

Y

it

= α

it

+ x’

it

β

it

+ u

it

i=1,…,N, t=1,…,T,

Where Yit is a scalar dependent variable, x’it is a vector of independent variables, uit is the scalar

disturbance term, i indexes individuals in a cross section and t indexed time. Even though the model presented is representative for panel models, it is too general and “further restrictions need to be placed

- 20 -

on the extent to which αit and βit vary with I and y, and on the behavior of the error uit” (A. Colin

Cameron & Trivedi, 2005)

For the econometric analysis of panel data, it cannot be assumed that the observations are independently distributed across time. Therefore, advanced methods have been developed to analyze it. In the following section, the most commonly used methods for panel data will be introduced. As it will be explained, the estimation of the coefficients (

β)

will vary in accordance to some specific assumptions of the method and the particular objective of the analysis.

7.2 Pooled OLS

Pooled ordinal least-squares or pooled OLS estimator allows for the purpose of this research to examine for average earning differences between sectors based on many factors such as individual or job-specific characteristics. The assumptions are very similar to more simple cross sectional analysis, but in addition, how relationships change over time can be studied by including dummy variables. Micro-econometric methods emphasize correlation over time for a given individual, with independence over individual observations. For SOEP 2010, being a country panel data, it is OLS estimate standard errors be corrected for clustering on the individual. Otherwise, default standard errors will assume that the regression errors are independent and identically distributed, which will mislead to inconsistent estimates of β. Moreover, the main disadvantage of pooled OLS method is that one cannot control for unobserved heterogeneity that is likely to affect earnings

As part of the initial analysis of the German working population, pooled OLS analysis is replicated in accordance to Hamilton (2000). Motivated by the assumptions previously discussed, the following income equation represents is presented to estimate the earning differentials between job sectors.

y

it

= α+ x’

it

β + e

it

(1)

where the function regresses “Y” hourly earnings for “i” individual in a “t” period of time, on a set of personal and family characteristics “x” and an error term “

e

it

”

. The variables used to control for the dependent variable income are explained in the analysis section. As explained before, consistency of OLS estimates requires that error term is uncorrelated with

x’

_it.

- 21 - 7.3 Quantile Regression

Moreover, alternative and more powerful methods will be used to compare the estimates. Quantile regression offers a robust and therefore efficient alternative to least squares estimations (Alexander, 2011). This regression offers the possibility to investigate how covariate effects influence the scale, location and possibly the shape of the earning distribution across sectors. Academics argue that the use of quantile regression methods allow researchers to obtain a more complete picture about the relationship between the dependent variable (hourly earnings) and the repressors x at different points in the conditional distribution of y. In this way, quantile regression methods can identify more precise effects which could be unnoticed while using a simple OLS regression. For instance, when there is a highly non-normal distribution (like figure 1 shows), OLS would be inefficient as it is sensitive to the presence of outliers. With QR, researchers have the possibility to calculate the least absolute-deviation regression, more commonly named as median regression, which is more robust. Moreover, QR differs from OLS as it does not require the existence of conditional mean for consistency of the estimates. In general, the standard linear conditional quantile function is the following:

Q

q

(y

i

|x

i

)=x

i

’β

q

Despite the advantages of the quantile regression over the OLS regression, it also presents some of the limitations. The most important limitation of these models is that it requires the assumption that the unobserved heterogeneity (α) is uncorrelated with the regressors in order to obtain consistent estimations of the β (Wooldridge, 2012) just as OLS does.

7.4 Fixed-effects

A more powerful model is the fixed-effect (FE). The FE model most importantly captures the unobservable heterogeneity (i.e. an individual ability that affects income/wages) but also it overcomes the limitations of the two previously presented models as the time-invariant component of the error term (αi) is permitted to be correlated with the regressors (x) of the model (Adrian Colin Cameron & Trivedi, 2009). In other words, the fixed-effects estimates will control for possible endogeneity of occupational choices in the earnings equation. Moreover, it assumes that the regressors’ xit are

- 22 -

uncorrelated with the idiosyncratic error (eit). The individual-specific-effect model for the dependent variable Yit t is the following.

y

it

= α

i

+ x’

it

β + e

it

(2)

Pooled OLS estimation assumes that [αi|xit] = 0 to ensure that self-selection into employment does not

follow on the basis of unobserved characteristics. Due to the nature of the research question, this assumption is very likely to fail and the OLS estimates become biased. By running a fixed effect model, it will be possible to estimate the relationship between earnings and belonging to a certain job sector in Germany. Moreover, the estimations of the regressors’ coefficients will explain certain amount of variation in earnings. From what is left over from thee estimated variation, fixed effects can control for individual-specific effects.

The main earnings effect gained from a fixed-effects model indicates the mean difference between the earnings from paid employment and self-employment of an individual who switches from one sector to another. Therefore, the fixed-effect model will provide evidence for theory of self-selection and earnings differentials. A limitation of the fixed-effects estimator is that the time-invariant variables are dropped out from the model and their coefficients cannot be identified. However, the present research does not present that inconvenient as time-invariant variables like gender are not needed to be estimated as the analysis focuses only males.

For the next section, the main interest is to provide empirical evidence supporting the existing differences in the earning distribution between self-employed and wage workers. First, an overview of the most recent descriptive statistics for the panel dataset will be presented. Second, in order to measure the effect of the regressors over hourly earnings, a pooled OLS regression will be implemented. This analysis will be presented by separating each of the job s. Third, to complement the OLS estimations, a quantile regression will be presented for both, self and paid individuals. The objective is to construct earning profiles as Hamilton (2000) did and provide evidence that can support or reject the theoretical models of investment, learning, superstar and compensating; behind the earning differentials. Last but not least, job sector will be combined in a las equation with the specified regressors in a fixed-effects regression. The regression will reveal evidence associated to earnings increase (or decrease) from switching from one sector to another. Moreover, fixed effects will control for possible endogeneity.

- 23 -

8. Results

8.1 Descriptive statistic

As mentioned before, it is not possible to provide summarizing statistic table for entire 1984-2010 study as the panel is unbalanced and making so could biased the result. Table 2 gives an insight of what the differences on the hourly income earnings look like for wage workers and self-employed. First, median measure was used as it takes into consideration the entire sample. In this case it is clear that the mean for self-employed is about 30% greater with respect to those paid workers. It is an interesting difference; however, the fact that it is a mean measurement should be flag precaution in making any final conclusions. Superstar entrepreneurs are present in the analyzed sample and might be the detonator for the substantial increase in the measurement. Second insight, the standard deviation is much more prominent for the self-employed. This result infers that the distribution on the earnings in skewed and confirms that a mean analysis might not be enough to make conclusions. Despite the fact that the standard deviation is larger for the self-employed, the one for wage workers is not small. It accounts for more than 50% earnings difference around the mean. A contrasting result is seen in the 25th percentile as the income for wage workers appears to be bigger. The earning difference is important between sectors, but is only representative for the people who earn less. And observation can be made in this point as in the determination of the sample, people who earned less than 1€ were filtered out of the sample. As a consequence, the results might shift down a bit but it does not represent a limitation for the purpose if this paper.

TABLE 2.A

Summary Statistics: Hourly Self-Employment Earnings and Wages for year 2010

Hourly Earning Measure

Statistic Wage Worker (1) Self-Employed (2)

Mean 17.22 € 21.00 € Standard deviation 9.33 € 18.38 € 25th percentile 11.06 € 8.66€ 50th percentile (median) 15.41 € 15.59 € 75th percentile 21.25 € 27.24 € Observations 3,846 518 *Euros in prices of 2010.

- 24 -

The 50th percentile is the leading indicator as it is neither influenced by very small nor very big values. In this sense, it has an advantage over the mean as superstar observations do not affect it. As it can be seen, the median for self-employment is bigger than for paid workers. The difference is not significant (1.4% bigger), but it contrasts to results from previous studies realized in other countries like the US, where the median earnings indicator was considerably higher for wage workers. Moreover, if the mean and the median are compared, a more homogeneous earning behavior can be observed for wage workers. The upper percentile is also higher for self-employed. In this case the difference is substantial and similar to the one of the mean difference. For this wave, only 4,465 observations were captured with the characteristics required for the study. More tables with descriptive results can be found in the appendix.

Despite table 2 provides evidence of the earning distribution among job sector, it is important to clarify that these differences are not only based on the fact that individuals belong to one or another sector. As mentioned before, there are certain factors that influence the earning for self-employed and wage workers. These factors can be individual specific or job specific. Some of them can be observed and others not. Therefore, in the following section multivariate findings will be evaluated.

8.2 Empirical Results

8.2.1 Pooled OLS & Quantile Regression

Due to the fact that the distribution of the earnings is much skewed as seen in figure 1, it was necessary to run ordinary-least-squares and complement the analysis with a quantile regression for the equation (1) presented in the last chapter. Table 3 provides the parameter estimates with the inclusion of the previously discussed control variables. The table is divided into two panels. Panel A corresponds to wage workers and panel B to self-employed individuals. The importance of making the pooled OLS and quantile regression into two panels it to observe the influence of the regressors over the earnings of each of the job sectors. Equation (1) is supplied with a quadratic function for potential labor market experience. The intention of having a quadratic function is to capture possible diminishing marginal effect. On the other hand, age was not included in the equation as it is highly correlated with work experiences, while this last constructed is based on the age of the individual.

- 25 -

Moreover, productivity across individuals can be determined by regressors such as the number of years of education that the individual received. Several dummies were included into the model. For instance, marital status of the individual, which was originally a categorical variable, was decomposed into 6 different dummies as described previously. The variable year of the survey was also included as time dummy variables. Last but not least, a dummy was set for disability status of the individual where “1” equals disabled and “0” not disabled. Note again that race is not taken into consideration. A small portion of the coefficients on table 4 are shown, complete tables are available at the appendix. No comparison among earning measures (Net profit, Draw, EAD) were able to be done as in the case of Hamilton (2000). However, it is interesting to see the differences across sectors and between regressions.

As expected, pooled OLS coefficients differ importantly from those of the quantile regression, even though for the ones of the median. Coefficients also vary across quantiles. For instance, the upper quantile coefficients are higher than the median, but smaller than the mean. Moreover, clustered standard errors also vary between quantiles, being these higher for the upper quantiles but smaller than the mean. Additionally, clustered standard errors are much higher for self-employed matching estimations previously taken by other authors (Hamilton, 2000; Kawaguchi, 2002; Praag et al., 2009). All of these differences are proven to be a consequence of the skewness of the earning distribution in which highly paid workers from each sector play an important role.

Having a closer look on the coefficients, it is noticeably that the highly statistical significant regressor WK_EXP (potential work experience) has a much greater impact at the upper quantile of hourly earnings for both, self-employed and wage workers. By the median, wage worker earnings seem to be more significantly affected by this variable, however. Moreover, a negative coefficient for potential work experience indicate a concave function and represents a diminishing marginal effect for wage workers at all levels of the distribution and for self-employed at least for the median and upper quantile.

- 26 - VARIABLE OLS QR_25 QR_50 QR_75 CONSTANT -17.90*** 4.302*** 4.133*** 1.985* (3.254) (.393) (.501) (.805) WK_EXP 0.410** 0.0102 0.0741** 0.198*** (.153) (.019) (.027) (.043) WK_EXPsq -0.00398 -0.000343 -0.00151** -0.00357*** (.003) (.) (.) (.001) 1.MARRIED 0 0 0 0 (BASE) 2.SINGLE -1.845 -0.201 -0.155 -0.141 (1.144) (.123) (.184) (.302) 3.DIVORCED -1.863 -0.439 0.216 -0.898 (2.676) (.237) (.677) (.996) 4.SEPARATED -1.737 0.117 0.0751 -0.333 (1.098) (.128) (.209) (.34) 5.WIDOWED 1.342 -0.0622 -0.167 0.204 (2.486) (.242) (.314) (.508) 6.OTHER -8.749* -0.636 -1.222 -3.714* (4.323) (.512) (1.367) (1.529) YEDU 1.802*** 0.0205 0.110*** 0.325*** (.151) (.02) (.026) (.041) DIS_STAT 0.143 0.0664 0.286 -0.101 (1.616) (.245) (.357) (.513) ~1984 0 0 0 0 (BASE) ~1990 2.313* -0.104 0.332 1.375*** (.956) (.217) (.261) (.386) ~2000 2.681*** 0.0402 0.869*** 1.728*** (.737) (.182) (.23) (.334) ~2010 6.389*** 0.153 1.298*** 2.649*** (.969) (.215) (.266) (.387) N 11621 2919 5811 8710 R-sq 0.121 0.017 0.04 0.078 * p<0.05, ** p<0.01, *** p<0.001

Parameter Estimates from Hourly Earning Regressions Regression

A.Dependent Variable: Earnings SE TABLE 3

- 27 - VARIABLE OLS QR_25 QR_50 QR_75 CONSTANT -8.318*** 3.521*** 3.959*** 1.827*** (.346) (.125) (.13) (.163) WK_EXP 0.386*** 0.200*** 0.230*** 0.294*** (.015) (.006) (.007) (.008) WK_EXPsq -0.00526*** -0.00353*** -0.00393*** -0.00487*** (.0003) (.0001) (.0001) (.0002) 1.MARRIED 0 0 0 0 (BASE) 2.SINGLE -1.416*** -0.410*** -0.570*** -0.742*** (.114) (.045) (.05) (.064) 3.DIVORCED -1.003 0.115 -0.173 -0.534 (.672) (.141) (.185) (.332) 4.SEPARATED -1.337*** -0.244*** -0.420*** -0.615*** (.188) (.061) (.075) (.101) 5.WIDOWED -0.229 -0.0239 -0.243** -0.285** (.269) (.083) (.086) (.109) 6.OTHER 0.263 0.251* 0.128 -0.259 (.309) (.11) (.122) (.148) YEDU 1.077*** 0.0794*** 0.0894*** 0.261*** (.024) (.009) (.009) (.011) DIS_STAT -0.371* -0.212* 0.0345 0.171 (.182) (.101) (.085) (.104) ~1984 0 0 0 0 (BASE) ~1990 1.851*** 0.212** 1.033*** 1.453*** (.129) (.066) (.052) (.055) ~2000 3.732*** -0.0108 1.170*** 2.608*** (.103) (.064) (.058) (.061) ~2010 6.370*** -0.0846 0.942*** 2.720*** (.148) (.088) (.08) (.083) N 104043 25958 52018 77918 R-sq 0.359 0.237 0.19 0.254

Standard errors in parentheses. Controlled for all year waves * p<0.05, ** p<0.01, *** p<0.001

TABLE 3 (Continuation)

Parameter Estimates from Hourly Earning Regressions Regression

- 28 -

In the case of the performance variables, it seems that for MAR_STAT (marital status), no earnings difference appeared to exist between married and single, divorced and separated entrepreneurs. Significant changes are more evident for wage workers where most of the categories had a negative coefficient under the base category. Thereafter, being married is the most profitable status. The coefficients for years of education are highly statistical significant for both job sectors. Despite the lower quartile for self-employment earnings, the regressor has an important effect. These coefficients are aligned with the evidence of other studies which conclude that education has and important impact on the earnings distribution (Praag et al., 2009). As Hamilton (2000) found, less educated German paid workers have a higher earning penalty than for their self-employed counterparts.

The effect on time is controlled by the survey year variable. In this section, table 3 shows four years (including the base year), which are enough to picture the tendency of the earning over the 27 years of the panel. As it is observed, coefficients are positive and increase year by year. A more detailed table with the coefficients of the 27 years can be found in the appendix. Compared to the work of Hamilton (2000) controlling for such a long period of time provides from a better understanding of the earnings between sectors and among time. For instance, the OLS coefficient shows that self-employment gains from hourly earnings were higher from 1984 to 1990 but then a sharp positive growth occurs for wage workers as the coefficient from 1990 to 2000 increases by a 100%. That mean increase can be attributed to highly paid workers as the coefficient increase for the median is not as high as it is for the upper quantiles.

8.2.2 Fixed-Effects

Before running the last regression it is important to verify which model is the appropriate one to use in the panel to obtain consistent estimates, fixed-effects or random-effects. For individual effects models, the main issue is whether the individual effect is correlated with regressors. The most popular test to know whether to use between FE and RE is the Hausman test. This test compares the estimable coefficients of time varying regressors and outputs evidence which model to use. However, the test presents a limitation as it requires the RE estimator to be fully efficient, and that is not likely to happen as there might be some unobserved time-variant individual characteristics that are serially correlated and therefore affecting earning levels. Therefore, a robust version of the Hausman test had to be performed, in which cluster standard errors could be assigned to the regressions.

- 29 -

The first step to accomplish the robust version of the Hausman test is by running a random-effects regression. Results from the RE estimation can found in the appendix. Table 7 provides the output of the test.

TABLE 4 ROBUST HAUSMAN TEST

Test of over-identifying restrictions: fixed vs random effects Cross-section time-series model: xtreg re robust cluster(ID) Sargan-Hansen statistic: 46.351 Chi-sq(6) P-value = 0.0000

As it can be noticed by looking at the p-value, the test strongly rejects the null hypothesis that random-effects provide consistent estimates. The estimates of RE might not be consistent as the provided regressors have a level of correlation with the individual-specific effects (αi). An example of it is an ability that an individual has, which does not change over time, but it is different from one individual to the next. Consequently, the test concludes that random-effect is not the appropriate model and therefore the fix-effect can be used with certainty to provide the estimates required to answer the research question.

Same control variables were used as previous regression. However, the constructed variable “potential work experience” had to be limited given the potential for multicollinearity with “year of survey”. According to Wooldridge (2012) “When we include a full set of year dummies—that is, year dummies for all years but the first—we cannot estimate the effect of any variable whose change across time is constant ”. As for the data of this panel “potential working experience” was constructed with age and education variables. This means that for each year that an individual grows, the “potential work experience” will also grow in a year unit. Moreover, the presence of αi accounts for differences across

individuals in their years of potential work experience in the initial time period. However, as the potential work experience increases in the same proportion for everyone, the effect of one year of additional potential work experience will not be able to be distinguished from the aggregated time effect (Wooldridge, 2012). For this reason, accounting for “years of education” (which is time-varying in this panel) and controlling time with “years of survey” would lead to a parsimonious model without multicollinearity issues.

- 30 -

TABLE 5 FIXED EFFECTS VS OLS ESTIMATES Dependent variable: Hourly Earnings

VARIABLE FE OLS CONSTANT -0.36 -9.203*** (1.331) (.441) SELF -EMPLOYED 1.346*** 1.609*** (.367) (.346) YEDU 0.639*** 1.156*** (.109) (.028) 2.SINGLE -1.428*** -1.433*** (.14) (.158) 3.DIVORCED -0.614 -1.142 (.357) (.675) 4.SEPARATED -0.766*** -1.362*** (.178) (.223) 5.WIDOWED -0.345 -0.0353 (.179) (.418) 6.OTHER 0.286 0.302 (.206) (.354) DIS_STAT -0.461* -0.332 (.181) (.205) ~1984 0 0 . . ~1990 2.512*** 1.881*** (.142) (.14) ~2000 7.247*** 3.566*** (.152) (.118) ~2010 11.35*** 6.361*** (.196) (.167) Observations 115670 115664 Within R2 0.155 . Overall R2 _{0.202 0.257} Between R2 0.207 .

Cluster standard errors in parentheses.

Controlled for all year waves. * p<0.05, ** p<0.01, *** p<0.001

- 31 -

Results from the estimation of fixed-effects regression with clustered standard errors can be found in Table 5. Notice that next to the FE estimation there is an extra column with OLS estimations. In this case, all the observations were set together in a single regression and complemented with a dummy variable for self-employment (same as for FE) in order to compare mean earnings between job sectors and to compare across estimation models. At first sight, the signs of the FE coefficients are consistent with those from the OLS and quantile regressions. The constant for the regression is not significant; however, this is not from major importance as it means that the constant term is not significantly different from zero.

The main explanatory variable for this regression is job type which has been renamed as “Self-employed”. This regressor is a dummy variable that equals “1” is the individual is self-employed or equals “0” if the individual is paid employee. As it was discussed before, it is the most interesting variable as it accounts for the process of self-selection for the different job sectors. This variable is highly significant and stands with a positive coefficient for self-employment. The results confirm that, after controlling for individual unobserved characteristics, the mean self-employed in Germany earns more than his salaried counterpart. The evidence contrasts to the predictions made by authors (e.g., Hamilton 2000;Kawaguchi 2002).

What it can be seen form comparing FE vs OLS is that fixed effects estimated coefficients are lower for all regressors. This results suggest that OLS coefficients where inflated by unobserved heterogeneity. On the other hand, the results from the marital status dummy variables are consistent with those of the OLS regression. Additionally, the disability status of the individual, being “1” for disabled, seems to have an impact on the hourly earnings when controlling for unobserved heterogeneity. Last but not least, the within R2 of the FE model explains 15.5% individual mean de-trended data, which seems to be a good indicator for fitness of the model.

Some scholars consider that using fixed-effects might have some drawback as the regression does not provide estimates of the magnitude the regressors have over earnings in each of the job sector. However, for the purpose of this paper, levels were captured by the pooled OLS previously run independently for self-employed and wage worker. The advantage for using fixed effect is not for estimating the impact of the regressors in each job sector but for measuring average differences between them when controlling for unobserved characteristics.

- 32 -

9. Discussion

9.1 Implications for theoretical models

The long tailed earning distribution and statistical evidence for self-employment is precisely predicted by the superstar theory. As it was observed from the density graph, there is a small portion of self-employed individuals who have enormous monetary remunerations. Interestingly, by analyzing in detail the observations from SOEP panel, high remunerated individuals resulted to be likely associated to medical and law advice professions. These professional titles are attached to high levels of education. Findings from quantile regression are supportive as the impact of education in the earning distribution is much higher for upper quantiles, and even more prominent for those who are self-employed. Based on the superstar theory, Hamilton (2000) suggested in his research that highly paid professionals should be taken out of the sample in order to make better mean predictions. Thereafter, a robustness check was realized in this paper dropping individual occupations similar to those described by Hamilton (2000); between them medical doctors, dentists and lawyers. An individual fixed effects model was run under the same conditions as for the original sample. Results from this test are in accordance from those of Praag et al.( 2009) as no significant existing difference was found between including or excluding these highly payed professionals in the panel. Detailed results can be found in table 6.

Matching and learning models state that individuals have sector-specific, time-invariant, unobserved

abilities which determine earning differentials. The matching model suggests that individuals will select which sector to enter based on in which they have relatively more advantage. Moreover, learning models argue that self-employees will have higher earnings as low ability entrepreneurs will drop the sector. By using an individual specific fixed effects regression, it was possible to control for unobserved heterogeneity and partially the effects of selection with respect to the enter self-employment. A positive coefficient in the “JOBTYPE” explanatory variable implies that the mean hourly earnings of self-employees are higher than the expected salary they would receive if they switch into paid-employment. Unobserved characteristics were proved to have an impact on earnings when the coefficients were compared to those of pooled OLS. Contradictory to the investment model, for the German population switching from paid employment into entrepreneurship seems not to be a bad idea. As the results from the fixed-effects showed in Table 3, shifting into self-employment has a significant