Does entrepreneurship pay?

(1)

Does Entrepreneurship Pay?

Student: Quentin Merelle

UvAID: 10827013/ VUnetID: 2566720

Contact information: quentinmerelle@gmail.com

University: Universiteit Van Amsterdam /Vrije Universiteit Amsterdam Program: MSc in Entrepreneurship

Supervised by Dr. Philipp Koellinger Date of completion: 29th June 2015

(2)

PREFACE

This MSc thesis was written to obtain the joint Entrepreneurship program of the Universiteit Van Amsterdam and Vrije Universiteit Amsterdam.

These last four months has been challenging but highly rewarding personally and professionally. This experience would never have enhanced my research and econometrics skills without the support and help of several individuals that I owe a tremendous gratitude.

That is why, I would like to thank my academic supervisor, Dr. Philipp Koellinger. He gave me the opportunity to make a research about such an interesting and exciting topic. His advices, knowledge, reactivity and organization gave me confidence in completing this thesis and, more importantly, to preserve challenges to handle a research project of quality.

I would like also to thank my teammates, Hugo Borja Bustamente, Umer Saqib, Patrick Suiker and Arne Weijling, with who I have been working on the same topic. Their comments and suggestions on my thesis were highly valuable.

Quentin Merelle 29th_{June 2015}

(3)

ABSTRACT

This paper analyses the potential different explanations of earnings differentials between self-employed and paid-employed, similarly as Barton H. Hamilton in 2000. The research takes place in the United States for elderly individuals between the age of 50 to 65 during the period 1992-2012. The empirical evidences show that most entrepreneur earnings and earnings growth are substantially lower than the one of wage worker. However, these differences are not valid anymore at a certain point of the population distribution. Education and work experience seem to have an increasing impact on earnings and could potentially fulfill this gap to some extent.

(4)

TABLE OF CONTENT

1. Introduction ... 7

1.1 Motivation ... 7

1.2 Research question ... 7

1.3 Relevance of the research ... 8

1.3.1 Academic relevance ... 8

1.3.2 Practical relevance ... 8

1.4 Outline of the thesis ... 9

2. Literature review ... 10

2.1 Being an entrepreneur or not: the economical choice ... 10

2.2 What evidence exists on the relative earnings of entrepreneurs compared to wage workers? ... 11

2.3 Which challenges do researchers have to face when trying to estimate the relative income of entrepreneurs compared to wage workers? ... 13

3. Methodology ... 15

4. Data ... 18

4.1 Data collection ... 18

4.2 Data description ... 19

4.3 Data analysis ... 21

4.3.1 The problematic zero earnings of self-‐employed, extreme outlier and number of observations for the panel study ... 21

4.3.2 Statistics descriptive ... 22

5. Empirical Results ... 25

5.1 Distribution of earnings ... 25

5.2 Pooled Ordinary Least Squares ... 28

5.3 Fixed effect ... 31

5.4 Quantile regression ... 33

6. Discussion & Conclusion ... 38

6.1 Empirical evidence compared to literature review ... 38

6.2 Limitations ... 39

6.3 Further research ... 40

6.4 Conclusion ... 41

7. Appendix ... 43

7.1 Empirical distribution of earnings measure of the panel dataset ... 43

7.2 Variable description and summary statistics with individuals observed at least one time ... 49

7.3 Variable description and summary statistics of individuals observed at least three times ... 56

7.4 Fixed effect without the work experience for significance of constant ... 61

(5)

TABLE OF FIGURES

Figure 1 - Empirical distribution, yearly earnings measure in 1992 ... 26

Figure 2 - Median income from 1992 until 2012 ... 28

Figure 7.1.1 - Empirical distribution, yearly earnings measure in 1994 ... 43

Figure 7.1.11 - Empirical distribution, yearly earnings measure from 1992 until 2012 ... 49

Table 1 – Impact of the number of observations per individuals filtering ... 22

Table 2 – Variable description and summary statistics for white male aged between 50 to 65 years old in 1992 ... 23

Table 3 – Summary statistics: Yearly self-employmed and wage workers earnings in 1992 ... 27

Table 4 – Pooled ordinatry least squares regression ... 30-31 Table 5 – Pooled ordinatry least squares regression with fixed effect ... 32-33 Table 6 – Quantile regression ... 35-36 Table 7.2.1 – Variable description and summary statistics for white male aged between 50 to 65 years old in 1994 ... 49-50 Table 7.2.2 – Variable description and summary statistics for white male aged between 50 to 65 years old in 1996 ... 50

Table 7.2.3 – Variable description and summary statistics for white male aged between 50 to 65 years old in 1998 ... 51

Table 7.2.4 – Variable description and summary statistics for white male aged between 50 to 65 years old in 2000 ... 51-52 Table 7.2.5 – Variable description and summary statistics for white male aged between 50 to 65 years old in 2002 ... 52

(6)

Table 7.2.6 – Variable description and summary statistics for white male aged between 50 to 65 years old in 2004 ... 52-53 Table 7.2.7 – Variable description and summary statistics for white male aged between 50 to 65 years old

in 2006 ... 53

Table 7.2.8 – Variable description and summary statistics for white male aged between 50 to 65 years old in 2008 ... 53-54 Table 7.2.9 – Variable description and summary statistics for white male aged between 50 to 65 years old in 2010 ... 54

Table 7.2.10 – Variable description and summary statistics for white male aged between 50 to 65 years old in 2012 ... 55

Table 7.2.11 – Variable description and summary statistics for white male aged between 50 to 65 years old from 1992 until 2012 ... 55

Table 7.3.1 – Summary statistics: Yearly self-employment and wages workers earnings in 1994 ... 56

Table 7.3.2 – Summary statistics: Yearly self-employment and wages workers earnings in 1996 ... 56-57 Table 7.3.3 – Summary statistics: Yearly self-employment and wages workers earnings in 1998 ... 57

Table 7.3.11 – Summary statistics: Yearly self-employment and wages workers earnings from 1992 until 2012 ... 61 Table 7.4.1 –Pooled ordinary least squares regression from 1992 until 2012 of white male aged between 50to 65 years old with fixed effect ... 61-62

(7)

1. I

NTRODUCTION

1.1 Motivation

Why become a self-employed? Apparently, the answer lies in the potential maximization utility that individuals can gain between being an entrepreneur or an employee. Researchers have found many factors that influence the career choice decision, but income seems to be one of the most important motivations for scholars (Campbell, 1992; Eisenhauer, 1995; Douglas & Shepherd, 2000).

The pecuniary benefits of self-employment are still a lively debate among scholars and practitioners. In fact, Hamilton (2000) wrote one of the most important researches about this fundamental entrepreneurship question: How well self-employed are doing compared to wage worker in term of monetary income? He found that there is a significant earnings gap between self-employment and wage workers income. On one hand, self-employed are earning relatively less than wage workers but on the other hand, within other parameters, it can be the contrary. Moreover, he also concluded that despite their lower earnings and lower earnings growth, individuals still enter and remain in the entrepreneurship sector.

Those differences between self and paid employment are the essence of this debate as the earning gap makes the decision making process more complicated when choosing an occupation. The current academic issue is still highly argued among scholars (Fairlie, 2005; Hamilton, 2000; McManus, 2000).

1.2 Research question

Barton H. Hamilton wrote in 2000 a pillar article in the field of entrepreneurship with an earnings comparison between self-employed and employee named “Does Entrepreneurship Pay? An empirical Analysis of the Returns of Self-Employment”. He studied the earnings distribution of American citizens in the 1970’s in the non-farmer sector. His objective was to determine “the extent to which the behaviour of workers choosing to enter or remain in self-employment can be explained by a variety of models of the labour market and entrepreneurship”. He showed interesting empirical results, as explained below, by focusing on several variables (e.g. working labour experience, marital status, education, health and race) and different earnings measurement (e.g: Net Profit, Equity adjusted Draw and Wage). His empirical evidences nourished the imagination of other scholars to analyse further this

(8)

interesting question. Therefore, my research question and hypothesis is much the same to the one of Hamilton “Does Entrepreneurship Pay?”. It covers why entrepreneurs would earn more/less than wage workers and under which circumstances. It is interesting also to know other potential contextual challenges that would complicate the understanding of this income gap.

1.3 Relevance of the research

1.3.1 Academic relevance

Based on the definitions and analysis conducted by Hamilton, described further below, my goal is to enrich the discussion about the monetary benefit of being self-employed or wage worker by enlarging the scope of the earnings comparison from a different dataset.

First, by using similar variables, sample design and types of analysis, my research fit the “requirements” to add significant scientific value and new perspectives of analysis in this field as it is easily comparable. In fact, my research will be used to be compared with other students` thesis which have similar settings as Hamilton also but with different samples. Our new incentives and reported results will be rewarding to deepen the actual empirical evidence.

Second, the time period and sample are different. It is highly interesting to compare the result of Hamilton in the 1980’s to mine from 1992 to 2012 of the same population (Americans) despite that my sample is different, to some extent, as described further below.

1.3.2 Practical relevance

Despite the fact that my dataset is focusing on a smaller and different aging scale of Americans, the elderly from 50 to 65 years, I believe that this paper will add value to the practical field. In fact, the sampling chosen are experienced individuals and from a dynamic region of the United States, Florida. Moreover, the panel dataset analysis, which is from 1992 until 2012, is witness of historical economic periods. The Internet crisis in 1998 and the economic crisis in 2008 have strongly affected the economy of the United States of America. From those events, there was a shift in the economy but also in individual’s occupation. New opportunities emerged in the labour market despite the rise of unemployment. The relationship between income and year of education and potential years of

(9)

experience for instance could be an interesting insight of the impact of human capital on the entrepreneurial income.

1.4 Outline of the thesis

After describing the theoretical framework, my methodology and the dataset, I explained my empirical result into four parts. The first one is dedicated to the earnings distribution between self-employed and wage workers so as to have a better overview and understanding of the main monetary differences that lies between these two types of occupations. Then, three regression analyses will be run to identify the earnings profiles and the relationship between my independent and dependant variables within my dataset: OLS, OLS with fixed effect and quantile regression. Finally, I purposed several discussion points and summarized the study through a conclusion.

(10)

2. L

ITERATURE REVIEW

As a short introduction part, I analyzed what drives individuals towards entrepreneurship from an economical perspective as it is interesting to have a clearer understanding of what are their expectations and needs. Then, I focused on the empirical earnings differences that lie between self and paid-employment. In this section, I also mentioned the potential reasons of those differences and challenges that scholars faced while analyzing this question.

2.1 Being an entrepreneur or not: the economical choice

Prior researches have been strongly focused on the psychological and sociological reasons of choosing an entrepreneurial career. However, since the 90’s, the scope and interest of analyzing this topic from the economical perspective increased (Campbell, 1992; Eisenhauer, 1995; Douglas & Shepherd, 2000). It appears that income is a crucial factor for the career choice decision-making process. It is as paramount and rewarding as the other entrepreneurial benefits such as being independent, passionate, career perspectives and so on (Douglas & Shepherd, 2002). Overall, choosing to be self-employed or not has been largely explained by scholars over time but, from an economical perspective, fewer evidence has been discovered so far.

Campbell (1992) introduced one of the first economic model about the decision making process of the entrepreneurial act. The essence of his model is to assume that the person will examine equally and subjectively the potential gain of being an entrepreneur and employee. He estimated the expected benefit of being an entrepreneur. It is determined by the difference between the probability of success and the average income of entrepreneur, and the probability of unemployment times the average incomes of employee (Campbell, 1992). The person will also take into consideration two other variables that are, by nature, correlated to the entrepreneurial act: attitude towards risk and uncertainty. However, it is important to remember that policy makers and economical regional development tools can highly influence the level of entrepreneurship and success. So, the results would be difficult to generalize for the entrepreneurship phenomenon.

Secondly, Eisenhaeur (1995) argued from another perspective the choice of being self-employed or not. From his study, he concluded that the probability of choosing to be self-self-employed depends positively on wealth, but negatively to income increases and insurance. Regarding the wealth, even Astebro (2010) explained that earnings expectations have little impact on the choice of

(11)

entrepreneurship. Regarding the income increase, even though there is a possibility of increasing it like wage workers, entrepreneurs seemed to do not attach a lot of importance. They mainly focus on increasing the value of their enterprises (Smith, Smith, & Bliss, 2011) and benefit from the other positive aspects of “being your own boss” (Frese & Gielnik, 2014). In fact, the high-risk of failure and uncertainty of new business venture (Smith et al., 2011) is not encouraging to perceive a revenue. “The run-of-the-mill startup is about 50% likely to fail within 6 years and is not very profitable” (Astebro, 2010, p.2). The expected cash flow is strongly advised to be invested in the enterprise valuation, not in the salary (Smith et al., 2011). Finally, when individuals chose to be an entrepreneur, it is not necessarily for the monetary benefit. Astebro (2010) also explained that the non-pecuniary effects might dominate the choice of becoming an entrepreneur as introduced above. As they are already satisfied within their working environment, they tend to rather focus on their enterprise than earnings.

2.2 What evidence exists on the relative earnings of entrepreneurs compared to

wage workers?

In practice, we tend to think that wage workers earn substantially more than entrepreneurs. From a scientific point of view, it is indeed similar. Hamilton (2000) and Astebro (2011) have been two of the most important scholars who clarified this point.

Hamilton concluded that self-employed have lower income and earnings growth than wage workers across most of the income distribution. In fact, even after 10 years of business, median entrepreneur would still earn less, on average 35%, than a predicted alternative paid job. If the job tenure would be zero, the pattern will be the same. Median entrepreneur would always earn less than the employee regardless the duration of their business experience (Hamilton, 2000). They would definitely earn more by starting a new job as an employee than remaining in the self-employment sector. In addition, it is also important to mention that there are quiet few self-employed that have negative earnings and therefore should not even entering into entrepreneurship (Hamilton, 2000). Moreover, the standard deviation of the income distribution is much higher for self-employed than for wage earners (Hamilton, 2000; Astebro, 2010). The distribution of self-employed is quite extended, some entrepreneurs are literally losing money whereas some are earning an extremely huge amount of money that even wage workers cannot even expect. They are so-called super-stars and can bias the

(12)

result to some extent. This will be further described below. It seems that being an entrepreneur is not encouraging also from the perspective of Astebro (2010) as they receive relatively lower wages than wage workers (mean and median), change jobs frequently or experience long unemployment periods than wage workers. He also concluded that 75% of all self-employed would have been better by not entering at all entrepreneurship. And according to Fairlie (2005), some evidences are still missing regarding the fact that business ownership provides actually a constant economic improvement.

However, sometimes the self-employed situation can be as attractive and rewarding than wage workers. In fact, individuals are willing to enter and remain self-employed despite the lower initial earnings and earnings growth than the alternative paid employment (Hamilton, 2000).

In 2000, Holtz-Eakin, Rosen, & Weathers found that over time low-income individuals tend to move towards self-employment for a better financial situation. Surprisingly, Fairlie (2005) found that even from disadvantaged families, young men business owner tend to have a higher income than wage workers. Hamilton (2000) explained that male entrepreneurs could experience greater initial earning growth in a new business than paid employment starting a new job. Even their potential wage of entrepreneurs is not significantly different from the one of paid employed (Hamilton, 2000) regardless the length of their previous entrepreneurial experience.

Moreover, even if the general tendency those entrepreneurs’ earnings are lower than wage workers, it tends to generally getting closer with wage workers’ earnings over time (Hamilton, 2000; Astebro, 2010). Indeed, after a longer entrepreneurial career, incomes are almost similar. It is for individuals who have at least 10 years tenure profile entrepreneurs (Astebro, 2010). And, it is also proven that at the 75th_{percentile of the distribution, self-employed tend to earn more on average than} wage worker whereas before they were earning less. Some of those individuals who earn an extremely higher income than employee, or even some of their entrepreneur fellows, are so-called “super stars”. But Astebro (2000) go further and analyze that the 99th percentile of self-employed earn 50% more than wage workers.

And, what about if there is a switch between the two status? If people switch, no matter the causes (financial regulation, new taxation rules, rising house prices, unemployment), for Astebro (2010), there is an average decline in earnings for people switching from wage work to entrepreneurship. But few do it because it requires a high human capital to have high return and face risks. For the other way around,

(13)

it is not the same. Despite their lower earnings, entrepreneurs remain in their sectors (Hamilton, 2000; Astebro, 2011).

2.3 Which challenges do researchers have to face when trying to estimate the

relative income of entrepreneurs compared to wage workers?

Measuring earnings between wage workers and entrepreneurs can be difficult because of the diversity of income tools measurement, honesty of the entrepreneurial reported income, outliers in the dataset and so on.

First, the most common issue is the under-reporting. This widely occurred income measurement is really problematic for scholars (Fairlie, 2005; Parker, 2006; Hamilton, 2000; Van Praag et al. 2009). Entrepreneurs can report their earnings as they want because they are totally responsible of their business, including financial matters. A majority of self-employed report their earnings as wage/salary income and not as business income (Fairlie, 2005). They can also evade income tax more easily than wage workers as they are responsible of their reporting of net profit for the tax authorities (Hamilton, 200; Parker, 2006).

Second, the study can be influenced negatively by high-income entrepreneurial superstars (Hamilton, 200; Parker, 2006; Astebro, 2011). It means that some entrepreneurs are earning an incredible huge amount of money compare to the rest of the group. The average earnings of the self-employed is therefore not representative of most business owners` earnings as super-stars increase the standard deviation and have a certain influence on the mean of income.

Last but not least, Hamilton (2000) explained a problem he encountered regarding the lack of information to determine if the difference between entrepreneur and employee earnings is due to true compensating differentials or difference in earnings growth across sectors. Others issues such as the job qualification, impact of fringe benefits, the non-response rate, parents education (Åstebro, 2011; Fairlie, 2005; Van Praag et al., 2009) are also proxies related to earnings that are not necessarily taken into consideration or at least not enough.

Surprisingly, “job spells” is an issue. The self-employment status for a panel study can be misleading. Indeed, the responses of job occupation can vary over time regarding the working labor

(14)

status, industry characteristics, coding protocol (Mc Manus, 2000). More specifically, some calculation of the coefficient between transitions between self-employment and wage for the same individual over time can be misinterpreted or misunderstood.

(15)

3. M

ETHODOLOGY

Based on Hamilton’s analytical framework and quantitative methodology, I replicated his analysis. Next to the summary statistics for the earnings measurement and the explanatory variables, Hamilton analysed the empirical distribution of the self and paid-employed in the year 1984 and 1985. He used three different income measurements.

1. Draw: It corresponds the most closely to wage of employee as it captures the equivalent consumption of business owners. It is less likely to be influenced by taxes as well.

2. Equity adjusted draw: It is the sum of draw and the change in their business equity over time. 3. Net profit: It is the difference between the revenue and the expense of the firm. But it does not

take into account how much of the profit is used for its own consumption.

Based on those tools, he concluded several interesting empirical evidence and patterns about the earnings differential between the two occupations and their income distribution.

My goal was to test and estimate the empirical evidences of the economic relationships and theories he found, but through different sampling settings that I described in the next section. I designed my statistical researches and results from a descriptive and explanatory perspective. It would ease the understanding of the econometrics methods used and, more importantly, its empirical evidences.

The statistical methods were computed from the statistical software: SPSS and STATA. Firstly, SPSS usage was dedicated to select variables and made the first statistic descriptive and analysis. Then, STATA was more convenient and handy to understand the estimation and effect of the parameters of the different regression models and at different percentiles of my empirical earnings distribution. It enables me to run three types of regression to diversify my research methods and find out others empirical evidence. Here below is an explanation of the interest in the regression equation for my panel study, except for replication purpose.

1. Pooled Ordinary Least Squares regression: Basically, the OLS is used to understand which are the most crucial relationship between the regressors and the dependent variables. Its goal is to analyse the importance of the impact of each variables on earnings. The magnitudes (ß) of the explanatory variables are then compared between the two regression models (one for self and

(16)

the other one for paid-employed) to see which variables has the biggest or lowest impact on earnings regarding individual’s occupation. The year of the interview and the marital status has been used as dummy variables in order to see the impact of each category. It was really useful to have this “yearly” category as the impact of time on earnings is really valuable and insightful for a panel study. For instance, we interpret if individuals at the year Y were earning on average more or less than individuals in the based year (1992). The standard errors of a regression are the estimated standard deviations of the unexplained variations in the explanatory variables. Regressors and errors could be correlated within a group for some reasons, so to ensure that the estimate of the pooled OLS give an appropriate inference, we shall clustering standard error as they are robust to heteroskedasticity (Cameron & Miller, 2013).

xtset hhdipn wageyear

reg iearn slfemp raedyrs workexpsq workexp i.mstat2 i.wageyear, vce(cluster hhidpn)

2. Pooled Ordinary Least Squares regression with a fixed effect: The fixed effect is used to analyse the impact of a variable that vary over time. It removes the effect of time-invariant characteristics on variables to assess the net effect of predictors on the outcome variables. (Torres-Reyna, 2007). This method ignore the between-individual variation and focus only on the within-individual variation (Allison, 2005) and the effect size is the same across individuals observations. So, the ability to control all stable characteristics of an individual reduces the potential disturbance so to more likely get unbiased estimates.

xtset hhidpn wageyear

xi: xtreg iearn slfemp raedyrs workexp i.mstat2 i.wageyear, fe

3. Quantile regression: In order to study the effect of the explanatory variables on earnings over the distribution, the quantile regression was the most suitable analysis. It is use to compare the earnings profiles over the distribution (25th, 50th and 75th percentiles) of self and paid-employed. It is the magnitudes of the coefficients regressors that are interpret to analyse the impact on earnings along the distribution. It is also possible to establish some relationships. For instance, if the coefficients grow over the distribution and the result are significant, it is proved that the effect of the regressors on earnings increases for individuals in the upper quantile. It is also

(17)

interesting to compare the coefficients of those three quantiles profiles to the ones of the OLS to understand the differences between what the overall regression shows (OLS) and what the more precise regression shows as well (distribution/quantiles).

xtile QPE=iearn if slfemp==0, nq(4) xtile QSE=iearn if slfemp==1, nq(4)

The following regress have been run also for QPE

regress iearn workexp workexpsq raedyrs i.mstat2 i.wageyear if slfemp==1 , vce(cluster hhidpn) reg iearn workexp workexpsq raedyrs i.mstat2 i.wageyear if slfemp==1 & QSE==1, vce(cluster hhidpn) reg iearn workexp workexpsq raedyrs i.mstat2 i.wageyear if slfemp==1 & QSE<3, vce(cluster hhidpn) reg iearn workexp workexpsq raedyrs i.mstat2 i.wageyear if slfemp==1 & QSE<4, vce(cluster hhidpn)

(18)

4. D

ATA

4.1 Data collection

The data used for this study are retrieved from the University of Michigan Health and Retirement Study. It is a national panel that surveyed every two years over 20,000 Americans and their spouses aged over 50 years old since 1992. The expected sample for my analysis is white males aged between 50 to 65 years old that were either self or paid employed for the following reasons:

1. Gender: For Hamilton article replication purpose, I also focus on male. 2. Race: Defined by my supervisor.

3. Age: It is more valuable to have people who are potentially working and not retired. Therefore, I limited my longitudinal study at the working labour limit age: 65 years old although the panel survey also incorporated a paramount number of older individuals. However, because of potential aging issues such as health or others reasons unexplained by HRS, some individuals can get retired before 65 years old or can even be out of the labour working market for any reasons related to aging or not. Thus, I used a labour working status to filter my dataset as described below.

4. Occupational status: In order to run a comparison analysis between the two types of individuals, I had to differentiate within my sample, individuals that were into the self-employment sector or the paid-employed sector. However, I could not determine if individuals were wage workers but only if they were self-employed or not self-employed. Therefore, the labour working status filter enable me to focus on “not self-employed individuals that were working full-time”, so hypothetically, paid-employed.

By doing so, I divided the number of individuals surveyed by ten from my original dataset. From 37, 319 individuals (including respondent, spouses and household) to 12,504 due to respondent focus, race and gender variables. Then, as I only had the birth year variable from the HRS and the interview period started in 1992 and finished in 2012, I focused on the “potential” 50-65 years old individuals. Those individuals were born between 1927 (=1992-651) and 1962 (=2012-502). In the end, I ended to 9,044 due to the birth year. But by framing my dataset in this way, I had some unnecessary individual’s

1_{The oldest possible individuals suitable for my dataset would be interviewed in 1992 aged of 65 and} therefore born in 1927. 2 The youngest possible individuals for my dataset would be interviewed in 2002 aged of 50 and therefore born in 1962.

(19)

observations for my panel study. An individual λ born in 1962 but interviewed in 1992 was aged of 30 years old which is not useful for my analysis. There were no data for this individual λ at the year 1992 (only missing value as it does not fulfil the HRS requirements) but the observation was still present. However, 20 years later (in 2012), this individual would be suitable in my dataset as he would be 50 years old. So, I created the “age” variable to increase the precision and reliability of my dataset and retrieved the individuals aged below 50 and upper 65 years. The number of individuals was the same but the number observations have been divided by two: from 99,484 to 46,077.

Finally the labour working status had also a certain influence on my dataset. By taking only into consideration the working full-time individuals, I had 4,399 individuals with 11,531 observations for the eleven wages of the period 1992-2012.

Nevertheless, due to the high percentage of missing values for the occupational variables coded as “whether self-employed or not”, the number of individuals surveyed has decreased by 54.6% ending up to 2,845 including 24.5% of self-employed and 75.5% of paid-employed and the number of observations plunged from 11,531 to 5,864. Similarly, I had to remove 14 individuals for a total of 17 observations because of their missing number of years of education.

4.2 Data description

In order to have an insightful statistic descriptive and regression equation, I used the earnings as dependent variables; educations years, occupational status and age as explanatory variables; and marital status and year of interview as dummy variables. Here below, their descriptions:

1. iearn refers to the annual earnings of the individual. Fortunately, these variables were similar for both self and paid-employed. I could not determine the hourly earnings because 80% of the weekly hours were missing on average for the eleven wages.

2. slfemp enable me to determine if individuals were either “self-employed” or “not self-employed”. However, the meaning of “not self-employed” was not precise because it could refer to unemployed, disabled, partly-working, etc. I added another controlling variables described below:

o Labour working status could categorize my two groups of individuals (self and not self-employed) based on their labour working status: working full-time, unemployed, disabled,

(20)

etc. I focused only on individual working full-time. It enables me to suppose that the not self-employed individuals who were working full-time were wage worker.

3. mstat is the marital status. It is an important explanatory variable in Hamilton study. The HRS dataset computed the individuals through eight different possible marital status so I created a more straight forward variable:

o mstat2 has been created to ease the interpretation of the regression model. People were simply categorized if they were either married with spouse present or not married.

4. raedyrs is the years of education of the respondent. It represents how many years the individuals has followed the education system. From zero until seventeen years, the dataset was quiet diverse. However, if individuals had more than 17 years of education, they were reported as a person with 17 years of education. To give a better insight of the American education system, here is a brief descriptive of the national education programmes based on the theoretical starting age from the International Standard Classification of Education (Unesco, 2007). From 5-7years old (Primary education); 11-13 years old (Middle education); 14-17 years old (high school); 18-30 years old (Bachelor); 22-30 years old (Master); 24-32 years old (PhD).

o raedyrshigher13 represent the number of people who have more or less than the average years of education. The variable was created based on the mean of education’s years which was thirteen.

o raedyrshigher17 represent the number of people who has been or not to University level.

5. rbyear is the birth year of the respondent and has been used to determine individuals born between

1927(=1992-65) and 1962(=2012-50) in order to have only the 50-65 years old individuals interviewed between 1992 and 2012 as explained previous page. It also enables me to create more specific and comprehensible variables for my study.

o Age has been created by deducing the year of the interview by the birth year. The result was the age of the individuals at the year of the interview which was more valuable.

o Agesquared has been used to have better estimate of the regression analysis.

6. workexp is the potential labour work experience and was not available in the HRS dataset but has been created based on Hamilton formula (= age minus education minus six)

o workexpsq has been used to have better estimate of the regression analysis.

7. wageyear is a created variable that represent the year of the interview. Each variables in the HRS panel was specify first by the person interviewed (respondent, spouse and household) and then the

(21)

code of the wages. I created the variables of the eleven wages where wage 1 was in 1992, wage 2 in 1994 and so on until wage 11 in 2012.

4.3 Data analysis

4.3.1 The problematic zero earnings of self-employed, extreme outlier and number of observations for the panel study

The first descriptive statistic was not encouraging. Firstly, there were 56.8% of business owners who were earnings zero dollars, and several employees also surprisingly (3.28%). But, this is a common issue for researchers as entrepreneurs have the power to under-report their income. Thus, I decided to delete them as it would bias my future earnings distribution analysis and quantile regression, mainly for the 25th and 50th percentiles. In the end, I focused on individuals who were earning more than 1 dollar. My population of self-employed is low so my sample size has slightly reduced when I removed those zero earning individual observations. From 2,831 individuals, my dataset ended up to 2,524 persons with 4,889 observations (including 82% of paid-employed and 18% of self-employed).

Secondly, I had to remove an extreme outlier observation of an entrepreneur aged of 57 years old in 1998. Despite the fact he fits my study criteria, his income of 3.53 million dollars was 68 times more than the average of the panel, 51,703 dollars, compared to 3.2 times more for the second outliers. Hamilton defined them as “super-star” and this individual could not be as representative as the others super-stars for my future analysis. Even though, I believed that removing one individual would not change a lot the income mean, it actually has been decreased by 711.6$. The standard deviation also changed and has been divided by 1.3.

Thirdly, the number of observations per individual was unbalanced for a longitudinal study. Table 1 shows the impact of my last and paramount filtering process in terms of number of individuals and observations. It was necessary to create this table as the statistic descriptive and the empirical results are based from two “different” datasets (1-7 observations/individuals and 3-7 observations/individuals) for the several reasons explained further.

Out of the 2,524 individuals, 48.3% were observed only once, 26.4% observed twice and the maximum number of observations was seven. Even though, I had eleven wages it is not surprising to have a

(22)

limited number of observations of seven as my focus is on elderly people aged between 50 and 65 years old and the survey is every second year which reduce the 15 potential years of observation by two. As my thesis was a panel study, it would not be reliable to have 75% of my sample that have one or two observations. I therefore focus only on individuals who were observed at least three times. This manipulation has a huge impact on my database. It reduced the number of individuals observed by three-fourth ending up with only 638 individuals for a total of 2,335 observations.

TABLE 1–IMPACT OF THE NUMBER OF OBSERVATIONS PER INDIVIDUALS FILTERING

INDIVIDUALS OBSERVATIONS

Individuals types

All observations

(before the filter)

More than 3 observations

(after the filter)

All observations

(before the filter)

More than 3 observations (after filter) Paid-employed 2,074 536 4,271 2,138 Self-employed 450 102 617 197 Total 2,524 638 4,888 2,335 4.3.2 Statistics descriptive

My longitudinal analysis encountered another complication. The numbers of observations of my 638 individuals were still unbalanced. More precisely, 378 individuals are observed three times (59.2%), 143 observed four times (22.4%), 80 observed five times (12.5%), 30 observed six times (4.7%) and 7 observed seven times (1.1%). Thus, my panel statistic descriptive would not be anymore valuable to some extent. Indeed, the more the individual λ is observed, the more the means and others statistical measurements will tend towards his values. It would be more interesting to focus on only one year though but my sample size was small. It would end up with an analysis with only 358 individuals if I would take the most observed year for instance. In front of this dilemma, although I previously deleted them and they represented 75% of my dataset, I decided, only for this analysis of statistic descriptive, to include my individuals observed one and two times in my data to increase my sample size. To reduce the biased effect of those “additional individuals”, I focused on one unique year and the most observed wages: It is the first one, in 1992, which represent 16.81% of my dataset (822 individuals).

Table 2 here below describes the differences between business owners and wage workers based on five variables. Overall, between 1992 and 2012, the number of the observations meeting the criteria of

(23)

being self-employed was 12.6% (ref: Table 7.2.11) but with the chosen year 1992 they represented almost 20%.

The potential labour market experience was not specified in the dataset, but Hamilton purposed in this study a formula that can estimate it: age minus years of education minus six. The difference is small between the two types of individuals but still significant for the year 1992, otherwise for the rest of the period it is not significant anymore. It is important to note the missing middle of the work experience here. Indeed, we agree that while the entrepreneur is developing his business, he is still gaining certain knowledge and experience in the labour market. This “post-business experience” is highly valuable but the entrepreneur cannot be considered as an active and full-time working entrepreneur within the labour workforce. He might be categorized as a self-employed, partly-working, partly-retired, unemployed or not at all in the labour workforce at all. In fact, those others non-active self-employment occupations represented 16.4% of my labour work status of white male aged between 50 and 65 years in the period of 1992-2012.

TABLE 2-VARIABLE DESCRIPTION AND SUMMARY STATISTICS FOR WHITE MALE AGED BETWEEN 50 TO 65 YEARS OLD IN 1992 MEANS T-TEST VARIABLE NAME DESCRIPTION Paid Employed Self-employed Differences between means workexp Potential labour market experience = age – education – 6 36.70 35.81 .88*

raedyrs years of education 12.72 13.60 -.88**

raedyrshigher13 13 and more years

of education .44 .58 -.13**

raedyrshigher17 17 and more years

of education .14 .20 -.06*

mstat2 Married, spouse

present .83 .83 0.0086

observations 662 160

(24)

Regarding the education, there are significant differences between entrepreneur and wage worker. Overall, it is clear that they tend to be more educated than employee. More precisely, there are 33% more of self-employed that study more than 13 years compared to paid-employed for a mean differential of -.13. The difference is less marked when we look at individuals who go to University but still present (-.06). These means differences between the two groups with 17 or more years of education fluctuates over the year (Table 7.2.1 to Table 7.2.10) so there is not enough evidence to conclude similar pattern for the rest of the period 1992-2012. Interestingly, the marital status does not make a difference if individuals are self or paid-employed in 1992 and neither for all the other years of observations.

(25)

5. E

MPIRICAL

R

ESULTS

As required for a valuable longitudinal study and explained in previous section, I focused on individuals who have at least been observed three times over the period 1992-2012. The following parts discussed about the empirical distribution of earnings differential between self-employed and paid-employed. Then, I analysed the pooled Ordinary Least Squares regression with and without a fixed effect. Finally, I will continue the regression analysis per quantile and occupational status.

5.1 Distribution of earnings

Compared to Hamilton, I had only one measurement tool for the earnings of self and paid-employed. There is no net profit earnings, after or before tax as it is purely and simply the “earnings: income” variable. Due to the unbalanced observations and as mentioned before, I focused on one year and I chose the most observed year (1992) of the self-employed population as they are less represented in the dataset. This year is the fourth most observed for wage workers.

Figure 1 draws the empirical distribution of income between entrepreneur and wage worker in 1992. In order to have a clear figure, I focused on individuals earning maximum 400,000$. Even though, the Y-scale ranges the data as 1e-6 and not 1 or 0.5, the area under the curve remains 1. The two density plots are less precise but still relevant. In 1992, the distribution of self-employed is less skewed than the one for wage worker. Over the panel, although the unbalanced number of observations, the distribution is similar for both types of individuals (Figure 7.1.1 to 8.1.11). More precisely, for instance, entrepreneur at the 40th percentiles are earning as much as the wage worker of the 25th percentile (23k$). There is approximately 0.7% of entrepreneurs who are earnings around 50,000$ compared to 1.3% for wage worker.

(26)

FIGURE 1 -EMPIRICAL DISTRIBUTION, YEARLY EARNINGS MEASURE IN 1992

Table 3 summarizes the yearly income measurement between entrepreneur and wage worker in 1992. We can clearly see from the income mean that being an entrepreneur is extremely advantageous. Their income is 43.68% higher than the one of paid employed. Despite the lower skewedness on the empirical distribution figure, here above, the standard deviation is 2.6 times more than wage worker. This difference is mainly due to the high number of outliers in self-employment. Their incomes are so outstanding compared to the rest of the population. They tend to significantly deepen the difference between the mean earnings of wage workers and self-employed.

Regarding the distribution, the people who earn the less should rather be wage worker than self-employed. For the median percentile, the situation is still in favour of wage worker but the difference between the two occupations is two times less than before. But for the upper tail of the distribution, the advantage shifted towards the self-employed as they earn more than wage worker with a difference of almost 20,000$. This is one of the reasons (the extreme high earnings) that the self-employed distribution is more skewed. Overall, being an entrepreneur seems really attractive if we look at the

(27)

overall distribution mean, but when we look at it more closely at the distribution it is only at the 75th percentiles exactly of the distribution that entrepreneur are earning more than wage workers.

TABLE 3-SUMMARY STATISTICS:YEARLY SELF-EMPLOYMENT AND WAGES WORKERS EARNINGS IN 1992

EARNINGS MEASURE STATISTIC Income

Paid-employed Income Self-employed Mean 42,624.8 61,244.44 Standard deviation 33,156.85 86,326 25th percentile 23,000 12,900 50th percentile 36,000 31,600 75th percentile 52,000 68,000 observations 249 36

Similarly, over the panel study (reference to the appendix from Table 7.3.1 to 7.3.11), if we look at the mean of the distribution, individuals should still rather be entrepreneurs. But it is only more profitable than wage workers only at the 75th percentile. Figure 2 represents the income median for both individuals and we can clearly see the dominance of wage worker over entrepreneur. More precisely, compared to the slow and smooth growth of paid-employed income, the evolution of the entrepreneurial income is increasingly and dangerously fluctuating over years. From 2003 until 2007, the income of self-employed actually took over the one of paid-employed. Nevertheless, this might be due to the lower number of self-employed observed per year (max. 28) that cannot represent the average of the dataset itself.

(28)

FIGURE 2–MEDIAN INCOME FROM 1992 UNTIL 2012

5.2 Pooled Ordinary Least Squares

After a short overview description of the OLS models, I will deepen my comparison analysis. Table 3 indicates two regression models for paid and self-employed. For the regression model, I retrieved the age variable as it gave me a negative constant and they are highly correlated to the one of work experience (=age-education-six).

At a first glance, the gap between the coefficients of regression models of self-employed and paid-employed are outstanding. Most of the entrepreneur’s predictors have bigger impact on the regression equation than the one of paid-employed. Even though, there is a lack of precision, their constant is approximately three times higher than the one of wage worker. And, over the years, the coefficients of the yearly dummies are clearly unbalanced. So, broadly, at a first glance, being an entrepreneur seems to be a bit more risky and responsible than being a wage worker.

The regression equation coefficients are really insightful with three independent variables (education, work experience and marital status) for both self and paid employed.

(29)

First of all, due to the similar precision of their coefficients, we can clearly see that education and work experience of both individuals have a significant impact on earnings. It seems like the better educated they are, the more they earn. And one the other hand, the more experience they have, the less they earn, surprisingly. Instead of assuming that the effect of work experience on earnings is linear for all years of experience, I squared the variables in order to have a non-linear and more accurate relationship. And, in fact, the more experienced they have, the more the individuals earns; +319$ for entrepreneurs compared to +69$ for wage workers.

But the biggest difference lies in the power of those coefficients. Entrepreneur’s coefficient’s magnitude are incredibly more important than the one of employee (e.g: 4,400$ for the regression of paid-employed and 10,800$ for the regression of self-employed). The impact of education and work experience on earnings for self-employed is 2.45 times higher for the former and 4 times higher for the latter compared to wage workers. In addition, the constant coefficients proved us that, if all ß are equal to zero, being an entrepreneur could be definitely more interesting but the lack of precision for this coefficients do not enable me to conclude such evidence.

Last but not least, marital status seems also more meaningful and powerful for entrepreneur than employee. But his p-value is not significant enough compared to wage worker who, in the end, should rather be married to increase their salary of more than 9,000$.

Overall, regardless the lack of precision for the constant of self-employed and by focusing on the explanatory variables, we can conclude that self-employment sector is apparently more beneficial than paid-employment. Regarding the earnings growth, it does not seem really stable to be a self-employed. Despite the none significant of my p-values and the low number of observations per year (3 to 28), there is still a trend that shows how disrupted is their earnings growth over the panel. The year 1994, 1998, 2002 and 2010 have negative coefficients. Surprisingly, the year 2010 is significant2. This lack of significance and reliability do not prove us any pattern for earnings growth of entrepreneur. Whereas for wage workers, the significance and the growth of the yearly coefficients demonstrates that

2_{Entrepreneur in 2010 earn less than in 1992 (the year omitted for a reference based). But when having} a look at the size effect of this test, it is not really valuable as there are three observations in 2010 compared to twenty-eight in 1992 and it is possible that those three individuals earn actually less on average than the other ones.

(30)

their income can significantly increase over time. For instance, on average, individuals in 2004 are earning 11,000$ more than in 1992 and so on. It seems that the potential earnings growth is higher and even more stable for wage workers.

TABLE 4-POOLED ORDINARY LEAST SQUARES REGRESSION REGRESSION VARIABLE NAME PAID-EMPLOYED (N=2,138;R2₌_0.1293) SELF-EMPLOYED (N=197;R2₌_0.1636) Constant 95,486.15** (41,806.43) 334,071 (198,354.3) slfemp 0 (omitted) 0 (omitted) raedyrs 4,404.318** (620.87) 10,814.6** (2,841.03) workexp -5,849.466** (2,018.49) -24,048.87* (10,018.01) workexpsq 69.95** (25.26) 319.41* (127.40) mstat2 9,267.555** (3,164.43) 27,788.18 (14,764.98) Year 1994 4,912.148* (2,140.18) -12,474.08 (15,164.98) Year 1996 5,127.66* (2,512.17) 7,956.81 (25,300.3) Year 1998 6,611.48* (2,697.97) -10,293.83 (16,169.8) Year 2000 6,924.27* (3,105.39) 13,701.59 (23,725.6) Year 2002 14,289.43** (4,337.10) -23,017.03 (21,537.44) Year 2004 11,018.16** (3,938.24) 17,487.8 (26,585.76) Year 2006 11,409.06* (5,013.97) 5,543.956 (39,107.45)

(31)

Year 2008 20,991.85** (5,728.02) 6,453.951 (18,974.61) Year 2010 21,078.51** (6,775.08) -65,914.15** (22,735.74) Year 2012 35,451.78** (10,880.56) 15,839.58 (54,652.13) *p>.05 **p>.01

5.3 Fixed effect

The regression model of this regression model is slightly different than the previous one. For replication and consistency purpose, I tried to use as much as possible the same model as the pooled OLS previously by keeping education years and work experience as independent variables, but not the squared ones. In fact no matter the different combination used, the constants were always negative and there were no significant p-values. Finally, I had a positive and valuable constant and two significant coefficients compared to no valuable statistics with the others potential regression. Last but not least, it is important to remember that the observations of panel dataset observations are unbalanced between individuals surveyed from 3 and 7 times. Here, the average observation is 3.7 which tend all value towards the least observed individuals.

Table 4 shows the impact of the fixed estimator on the pooled OLS regression for self-employed and wage worker. We can clearly see the potential benefit of work experience. The more the individuals have experience in the labour market, the more they earn (almost 1,000$/year). However, this is as valid for self-employed than wage worker. The difference between self-employed and paid-employed lies in their decision of career path. It seems like choosing to be an entrepreneur might be the right choice as it would increase the income by almost 4,000 dollars but the p-value is not precise enough to conclude this statement. The other coefficients were not significant. The yearly dummies do not show if there is a significant evolution of earnings over time.

I tried others models such as in Table 7.4.1 (appendix) which considerably increased the precision of some estimates, including the constant. Their constant will be +47,766.57$/year with this other regression (Table 7.4.1) compared to +20,729$ for the regression in Table 4. Unfortunately, this new

(32)

model still did not enable me to compare accurately how entrepreneurship can pay as the work experience variable has to be retrieved so I cannot compare with the others model. But this reflection demonstrate that if the effect of those independent variables would be the same (fixed), there would be less evidence to explain the potential benefit of the explanatory variables. On one hand, the number of educations years has been omitted due to collinearity. But, on the other hand, the marital status and the career choice seem to still have a similar but smaller impact the income of individuals than the pooled OLS. Therefore, it seems like removing the timing characteristic disrupt the previous initial evidence about education years (it is omitted) but “confirmed” significantly the relationship between earnings and work experience as in the pooled OLS regression.

TABLE 5-POOLED ORDINARY LEAST SQUARES REGRESSION WITH FIXED EFFECT REGRESSION

(N=2,335)

VARIABLE NAME Coefficients

constant 20,729.13 (11,219.5) slfemp 4,055.97 (4,222.75) raedyrs 0 (omitted) workexp 862.38** (293.62) mstat2 336.54 (4,842) Year 1994 -72.2548 (2,762.35) Year 1996 -2745.36 (2,830.35) Year 1998 -1723.66 (2,796.29) Year 2000 -2720.79 (3,135.04) Year 2002 -2934.14

(33)

(3,680.23) Year 2004 -6618.79 (3,738.37) Year 2006 -4099.731 (4,397.37) Year 2008 -457.41 (5,024.68) Year 2010 -5154.65 (5,716.95) Year 2012 0 (omitted) *p>.05 **p>.01

5.4 Quantile regression

The table 5 illustrates how the effect of the explanatory variables and dummies vary over quantiles and how the magnitudes of these effects of various quantiles differ from the OLS coefficient.

Generally, the estimates for all quantiles are smaller in magnitude than the OLS for both individuals. But, the OLS and quantiles regression coefficients for entrepreneur are stronger and with more extreme values - from the lower to upper quantiles - than the one of wage workers who have more concentrated coefficients along the distribution. This pattern reflects well the distribution of the entrepreneurial and employee earnings. However, it is important to remind that the table 4 do not show enough evidence, because of lack of significance, regarding the estimates of self-employed. Thus, some of the relationships explained below are not as relevant for this group as for the wage worker. For instance, it is surprising to see the coefficients of the .75 quantile below the one of the OLS whereas it should above the average, the OLS though. This might be due to the lack of precision of the model.

Over the distribution, the constant coefficients of self and paid-employed have a similar “growing pattern”. The coefficients increase until the 50th percentiles but then slightly decrease for the higher level of the distribution. So, it seems like it is more beneficial for individuals to be among the median earnings profiles than being among the “higher earnings group” (the 75th percentiles). But,

(34)

there is a biggest significant difference across quantiles when looking at their coefficients magnitude that are apparently higher and more significant for wage workers. We can clearly see that being a wage worker seems to be more advantageous. Indeed, self-employed coefficients impact are little lower than paid-employed at the .25 and .50 quantiles but, surprisingly, extremely lower at the .75 quantile.

Regarding education, wage workers who have one more years of education earn 461$ less for those with low income (at the 25% quantile) and earn 550$ more for those with higher income (at the 75% quantile). In other words, the effect of number of years of education increases for wage workers with higher earnings. Note that the number of education years is only positively influencing income at the upper quantile. The effects on earnings from the self-employment sectors seem bigger across all quantiles but there is a lack of precision to conclude this statement as well.

In addition, work experience is not anymore such an important determinant of the regression. Along the distribution, the potential labour market experience is indeed not sufficiently precise but marital status is. Actually, by comparing the regressors, it seems like marital status has also the largest effect on earnings among the others explanatory variables. And across quantiles, if people are married, they are generally earnings more by being either self or paid-employed. But, the effect of marital status on earnings seems bigger for entrepreneur than wage worker all over the distribution. So, no matter in which quantiles they are, individuals should rather be in the self-employment sector if they want to earn more when married.

However, if we don’t focus on what individuals chose or not as a career, the highest benefit of marital status they could retrieve from would be at the .50 quantile for entrepreneur and .75 for employee. In fact, this is where the magnitudes have the highest statistical significance in their respective distribution. On one hand, if the individual is an employee, the effect of marital status has significantly a stronger effect on earnings at the median and then upper quantiles3 whereas for entrepreneur is significantly powerful at the lower and then median quantiles4.

In addition, along the distribution, the .75 quantile is the most insightful. The wage worker earnings of the upper-quantile have been more “impacted” by regressors than the one of the

3_{There is no significant p-values at the lower quantiles although the increase of coefficients in the distribution.}

(35)

entrepreneur. In fact, although no variables have a significant effect on the earnings at the 75th percentile for self-employed, the constant, education and marital status have a significant impact on the earnings for employee.

Overall, although there is a low statistical significance of the self-employed regression, we can definitely see how they are dominated by wage worker to some extent and how difficult it is to understand the impact of the explanatory variables over the distribution for entrepreneur. There is no proof for self-employed to explain which variables has the biggest impact on the earnings differentials at the 75th percentiles, point at which entrepreneur earn more than wage worker according to the previous analysis. So, there is not really a “best earnings profiles” for them if only focusing at the coefficients of each quantiles. But, even though .75 quantile is still statistically speaking more powerful, we do not know how stronger is it compared to others quantile due to low precision. However, the earnings of wage worker are increasing over the distribution as expected from the empirical distribution. And education and marital status are constantly present over the distribution and have a significant and increasing impact on earnings.

TABLE 6- QUANTILE REGRESSION

REGRESSION

VARIABLE NAME OLS .25 .50 .75

A. Dependent variable: Earnings of Self-employed (N=197;R2=0.1636) Constant 334,071 (198,354.3) 17,211.24 (28,229.31) 39,410.35 (48,510.82) 9,395.60 (96,612.4) raedyrs 10,814.6** (2,841.03) -142.0636 (281.18) 333.1767 (504.16) 1,335.714 (1,128.42) workexp -24,048.87* (10,018.01) -215.2496 (1,348.58) -1,656.014 (2,468.68) -417.7407 (5,057.64) workexpsq 319.41* (127.40) -.5051959 (17.38) 20.88 (31.99) 8.76 (66.45) mstat2 27,788.18 (14,764.98) 3,475.416* (1,598.78) 4,945.08* (2,296.56) 5,708.021 (5,334.81) Year 1994 -12,474.08 -1,584.328 1,442.696 -599.2251

(36)

(15,164.98) (1,387.09) (3,324.11) (4,695.19) Year 1996 7,956.81 (25,300.3) -3,006.921 (2,364.85) -3,102.334 (4,144.20) -109.249 (6,921.88) Year 1998 -10,293.83 (16,169.8) -3,437.793 (2,479.97) -6,264.729 (3,922.19) -4,610.428 (7,795.63) Year 2000 13,701.59 (23,725.6) -259.946 (2,623.85) 57.48 (5,625.21) -1,095.686 (9,651.58) Year 2002 -23,017.03 (21,537.44) -4,717.538 (2,479.97) 1,079.745 (3,555.87) -5,771.152 (8,304.96) Year 2004 17,487.8 (26,585.76) -2,912.643 (2,733.23) -3,972.80 (4,703.01) 651.7018 (10,813.63) Year 2006 5,543.95 (39,107.45) -2,264.298 (2,430.34) -1,088.356 (5,540.95) -3,576.426 (12,935.35) Year 2008 6,453.95 (18,974.61) - 11,832.92 (8,002.35) 13,230.34 (15,069.07) Year 2010 -65,914.15** (22,735.74) -3,921.686 (2,411.35) -6,859.607 (4,430.53) -12,527.35 (17,476.18) Year 2012 15,839.58 (54,652.13) - -2,334.091 (3,202.96) 627.90 (17,948.24) B. Dependent variable: Earnings of Paid-employed

(N=2,138;R2=0.1293) Constant 95,486.15** (41,806.43) 30,196.08 (18,472.47) 72,355.17** (15,750.56) 65,781.65** (18,888.35) raedyrs 4,404.318** (620.87) -461.6896** (176.99) -350.8182 (206.89) 550.98* (262.27) workexp -5,849.466** (2,018.49) -73.51 (973.98) -1,672.60* (807.67) -1,644.24 (936.71) workexpsq 69.95** (25.26) -4.54 (12.23) 11.82 (10.29) 11.84 (11.86) mstat2 9,267.55** (3,164.43) 1,173.99 (1,061.78) 3,471.975** (1266.31) 4,485.611** (1,578.6) Year 1994 4,912.15* (2,140.18) 2,091.934* (967.37) 2,271.426* (925.92) 1,589.398 (1,031.39) Year 1996 5,127.66* 1,261.537 1,027.282 3,367.47*

(37)

(2,512.17) (940.82) (1,124.19) (1,315.10) Year 1998 6,611.48* (2,697.97) 547.8464 (1111.721) 809.34 (1,147.47) 3084.868* (1,334.56) Year 2000 6,924.27* (3,105.39) 586.0388 (1298.381) 2,127.435 (1,246.35) 2,939.906 (1,513.24) Year 2002 14,289.43** (4,337.10) 753.5086 (1512.496) 2,817.327 (1,533.79) 3,988.871* (1,806.45) Year 2004 11,018.16** (3,938.24) 767.3297 (1,540.69) 1,908.758 (1,517.10) 2,546.224 (1,793.56) Year 2006 11,409.06* (5,013.97) 276.9021 (1,789.90) 1,274.754 (1,829.27) 2,221.024 (2,156.38) Year 2008 20,991.85** (5,728.02) 1,279.10 (1,941.88) 1,209.49 (2,299.65) 7,935.56** (2,562.62) Year 2010 21,078.51** (6,775.08) -764.3361 (2,607.33) 1,462.854 (2,588.8) 5,479.897 (3,097.83) Year 2012 35,451.78** (10,880.56) -1,622.857 (2,712) 2,344.408 (3,032.65) 9,192.69* (3,581.55) *p>.05 **p>.01