SCHUMPETERIAN ENTREPRENEURSHIP

Transcription

1 Paper to be presented at the Summer Conference 2009 on CBS - Copenhagen Business School Solbjerg Plads 3 DK2000 Frederiksberg DENMARK, June 17-19, 2009 SCHUMPETERIAN ENTREPRENEURSHIP Atsushi Ohyama University of Illinois at Urbana-Champaign atoyama@igb.uiuc.edu Serguey Braguinsky Department of Social and Decision Sciences, Carnegie Mellon sbrag@andrew.cmu.edu Steven Klepper Department of Social and Decision Sciences, Carnegie Mellon sk3f@andrew.cmu.edu Abstract: Based on recent findings concerning the best performing startups, we develop a model of Schumpeterian entrepreneurship in which founders exploit ideas they learned through their employment. The model yields distinctive implications about how labor market experience and earnings at work influence the probability of a worker becoming an entrepreneur, earnings as an entrepreneur relative to paid work, and persistence in entrepreneurship. These implications are tested using data on the earnings of scientists and engineers, which are common founders of high growth startups. The sample is pared down to those that worked in and founded businesses related to their education in order to isolate the best candidates for Schumpeterian entrepreneurship. JEL - codes: M13, J31, -

2 Schumpeterian Entrepreneurship Abstract Based on recent findings concerning the best performing startups, we develop a model of Schumpeterian entrepreneurship in which founders exploit ideas they learned through their employment. The model yields distinctive implications about how labor market experience and earnings at work influence the probability of a worker becoming an entrepreneur, earnings as an entrepreneur relative to paid work, and persistence in entrepreneurship. These implications are tested using data on the earnings of scientists and engineers, which are common founders of high growth startups. The sample is pared down to those that worked in and founded businesses related to their education in order to isolate the best candidates for Schumpeterian entrepreneurship. JEL classification: M13, J31 [Key words: entrepreneurship, scientists and engineers, earnings differentials, self employment] [Running Title: Schumpeterian Entrepreneurship] This is an abridged version of the paper that has been cut to 10,000 words by the authors for the purpose of the DRUID conference review only. Please do not cite this version of the paper. The full version of the paper is available upon request.

3 Schumpeterian Entrepreneurship 1 Introduction Numerous visions have been articulated about the role of the entrepreneur in a capitalist economy. Perhaps the best known is Joseph Schumpeter s view of the entrepreneur in the Theory of Economic Development. Schumpeter s entrepreneur is an agent of change that is the source of his famous creative destruction. He introduces a new good or a new method of production, opens a new market or discovers a new source of supply, or carries out a new organization of an industry. He upsets the conventional way of doing things. When successful, he elicits widespread imitation. Such success presupposes a great surplus force over the everyday demand and is something peculiar and by its nature rare (Schumpeter [1949]). What is the impetus for this kind of Schumpeterian entrepreneurship? Popular theories of entrepreneurship feature the role of risk taking (Kihlstrom and Laffont [1979]), managerial ability (Lucas [1978]), wealth (Evans and Jovanovic [1989]), and preferences for the control, flexibility and other job attributes that come with being one s own boss (Hamilton [2000]) as the primary motivations for entrepreneurship. While all receive support from empirical investigations of the self employed and business owners, none maps into entrepreneurship with the singular impact envisioned by Schumpeter. Studies of the most successful business startups suggest they are quite different from the typical new business. They tend to be founded by talented individuals based on ideas they encountered through their prior employment (Bhidé [2000], Kaplan et al. [2005], Klepper and Thompson [2009]). Consistent with Schumpeter s vision, they seem to be more about ideas and talent and less about individual characteristics and preferences, with the ideas originating from work experience. We develop a model of such entrepreneurship in order to derive various testable implications that can be used to probe its importance and functioning. In the model, employees of incumbent organizations are assumed to continually receive ideas of unknown value that they could develop in their own firms. The return from developing an idea depends upon the talent of the employee as well as the value of the idea. Talent also conditions an employee s earnings as a worker, but yields a greater payoff when 1

4 combined with a valuable idea. Initially employees cannot differentiate the value of ideas. The average idea is not worth pursuing and so at first no ideas are developed. As they gain labor market experience, employees become better able to judge ideas and the best prospects are developed in startups. Firms are disproportionately founded by more talented individuals, who have a lower threshold in terms of the expected value of ideas for entrepreneurship to be profitable. Entrepreneurs learn from experience about the value of their ideas. If the expected return from continuing to pursue an idea falls below the wage they could earn from work then they exit entrepreneurship for paid work. The model yields distinctive implications about how labor market experience and earnings at work influence the probability of a worker becoming an entrepreneur, earnings as an entrepreneur relative to paid work, and persistence in entrepreneurship. We test the model using data on the employment and earnings of scientists and engineers, which are a common source of founders of the best startups. In order to pare down the sample to those most likely to be Schumpeterian entrepreneurs, we consider only individuals that work in and start businesses related to their education. Exploiting the panel nature of the dataset, we explore the factors that influence workers to start businesses, the determinants of the earnings of new business founders relative to their earnings as workers, and the factors that influence the hazard of entrepreneurs returning to paid work. With the exception of some of the findings for older workers, our tests provide strong support for our model. We attribute the departures for older workers to our imperfect ability to pare from the sample workers that choose entrepreneurship for defensive reasons. The paper is organized as follows. In Section 2, we review past findings regarding the self employed and business founders. In Section 3 we lay out our theoretical model of the Schumpeterian entrepreneur and derive various predictions. In Section 4, we discuss data and methods. In Section 5, we test the predictions of our model. In Section 6, we consider alternative explanations for our findings and discuss the implications of our findings for further research on Schumpeterian entrepreneurship. 2

5 2 Past Findings 2.1 General Findings Various patterns emerge from the literature on the self employed and business startups. Sectors where the number of self employed is greatest include, in order, construction, retail trade, professional services, business services, and real estate (Lazear [2005]). The percentage of individuals that are self employed rises with age until around age 45 and then levels off through age 60 or so, which is consistent with a constant hazard into and out of self employment. The self-employment rate then rises sharply in the age bracket (Evans and Leighton [1989]). The least and most educated tend to have the highest percentage of self-employed individuals (Blanchflower [2000]). Earnings of the self employed are less than the earnings of workers at the median and lower percentiles but greater at the higher percentiles, reflecting the lower median but greater variance and skewness in the distribution of the earnings of the self employed than workers (Hamilton [2000]). Self employed individuals report being more satisfied with their jobs than workers (Blanchflower [2000]). Various theories have been advanced to explain these patterns. Lucas [1978] envisions individuals sorting into business owners/leaders and workers according to their managerial skills, which could explain why self employment is more prevalent among the most educated. The most educated are also the highest earners and thus have more wealth to finance their own business and possibly a greater willingness to bear risk, both of which have been conjectured as factors contributing to entrepreneurship (Evans and Jovanovic [1989], Kihlstrom and Laffont [1979]). Many individuals are thought to become entrepreneurs to take advantage of the greater control and flexibility that comes with being one s own boss, which is consistent with the greater job satisfaction reported by entrepreneurs. This could also explain the willingness of individuals to become entrepreneurs despite the lower median earnings of entrepreneurs and the lower riskadjusted returns of owners of private versus publicly-traded businesses (Moskowitz and Vissing-Jorgensen [2002]). Analyses of the transition into entrepreneurship and the performance of entrepreneurs have provided support for each of these motivations, although not without 3

6 dissent. Factors conjectured to improve an individual s skills as an entrepreneur, including having self-employed parents (Lentz and Laband [1990], Lindh and Ohlsson [1996], Dunn and Holtz-Eakin [2002]), working in smaller firms (Sorensen [2007], Elfenbein, Hamilton, and Zenger [2008]), and having more diverse work and school experiences (Lazear [2005], Wagner [2006]), all raise the probability of becoming an entrepreneur. Wealthier individuals, including those receiving inheritances and lottery windfalls, are more likely to start businesses (Evans and Jovanovic [1989], Holtz-Eakin, Joulfaian, and Rosen [1994a,b], Lindh and Ohlsson [1996], Blanchflower and Oswald [1998]), with wealth also influencing the performance of the businesses (Holtz-Eakin, Joulfaian, and Rosen [1994a]). This is consistent with wealth easing liquidity constraints, but wealth could also be a proxy for human capital (Nanda [2008]), with human capital determining both the performance of a new business and the capital it can raise (Bates [1990], Cressy [1996]). Alternatively, wealth could be a proxy for risk or the ability to purchase the luxury benefits associated with owning one s own business (Hurst and Lusardi [2004]). Consistent with individuals choosing entrepreneurship based on the benefits it affords, those with a stronger internal locus of control and preference for being their own boss are more likely to become entrepreneurs (Evans and Leighton [1989], Burke, FitzRoy, and Nolan [2000]). 2.2 The Best Start-Ups Schumpeter s entrepreneur is certainly not the typical self-employed individual/business founder. Three recent studies concerning the best-performing startups are instructive about the best candidates for Schumpeterian entrepreneurship. Bhidé [2000] studies firms in 1989 that were on Inc magazine s top 500 list, which is a compilation of the fastest growing privately held firms in the U.S. Kaplan, Sensoy, and Strömberg [2005] study the business plans and descriptions at (on average) ages 2, 5, and 8 of VC backed firms that were able to go public and also all non-financial firms in 2004 that engaged in an IPO. Klepper and Thompson [2009] summarize various statistical regularities and case studies concerning businesses founded by employees of firms in the same industry (called intra-industry spinoffs), which consistently outperform other startup entrants. We synthesize from these studies two key themes. 4

7 First, the various groups of firms were concentrated in high-growth industries and generally exploited distinctive ideas, which their founders mainly developed through their prior employment in the same or related industry. In the Inc 500 firms studied by Bhidé [2000], the greatest number of firms came from computer-related areas, which experienced high growth. Approximately 70% of the firms replicated or modified an idea their founder had encountered in his previous employment, with 91% founding their firm within a year of the genesis of the idea. Many had to modify their ideas over time in order to succeed. In Klepper and Thompson [2009], case studies of the origins of the leading intra-industry spinoffs in autos, semiconductors and lasers, all high-growth industries in their time, revealed that their founders commonly exploited technical and organizational ideas that arose at their prior employer (in the same industry). In many instances, the impetus for the spinoffs was a disagreement in their parent firms over their ideas. Among the VC-backed startups studied in Kaplan et al. [2005], over half were in biotechnology and software/information technology, both high growth areas. In virtually every firm, there was little change in their strategic focus from age 2 onward whereas there was considerable turnover in management, reflecting the greater importance of ideas than people. Second, founders of each group of firms were invariably highly educated, accomplished individuals that had risen up the ranks of management but did not anticipate or perceive themselves as entrepreneurs and often were not the long-term leaders of their startups. Bhidé [2000] reported that 83% of the founders of the Inc 500 firms he studied had a college degree and 35% an advanced degree. Many did not perceive themselves as doing something especially risky or noteworthy, perhaps because over a third had been fired or left their job over a disagreement and many others continued to perform their old job while starting their firm. Among the leading intraindustry spinoffs studied by Klepper and Thompson [2009], many were founded by distinguished technical people who were also involved in management, often at high levels. The impetus for starting their firm was often a controversy in their employer over an idea. Many like Gordon Moore of Intel and Fairchild Semiconductor fame described themselves as unintentional, accidental entrepreneurs (Moore [1994]). Judging from the findings of Kaplan et al. [2005], over time founders receded in terms of the 5

8 management role they played in their firms. At the time of the business plan around age 2, the founder was the CEO in 66% of the firms, which declined to 58% by the time of their IPO (on average at age 5) and then 39% around three years later. We do not want to overstate the overlap in the findings of the three studies nor downplay the findings of the numerous studies that have tried to uncover the factors that lead to successful entrepreneurship. But the firms studied by Bhidé [2000], Kaplan et al. [2005], and Klepper and Thompson [2009] are certainly different from the average startup. They were founded mainly in high growth areas, by talented, highly educated individuals with management experience. They exploited ideas of uncertain value their founders had recently come across through their employment. The importance of ideas is certainly consistent with Schumpeter s conception of entrepreneurship and with the work of others as well (cf. Holmes and Schmitz [1990]). However, ideas have not been featured in mainstream theories of entrepreneurship (Shane and Venkataraman [2000] and Shane [2003]), perhaps because they are not central for the average startup. In our model of Schumpeterian entrepreneurship, though, the random generation of ideas of uncertain value drives the model. The prominence of high level people with management experience as founders but not necessarily managers of successful new firms suggests that management experience may be mainly important in evaluating new ideas, with talent key to exploiting the ideas. We will feature both of these ideas in our model of Schumpeterian entrepreneurship as well, to which we now turn. 3 Theory Our vision of a Schumpeterian entrepreneur, outlined in the previous section, is of a talented, formally educated individual who comes across a worthwhile idea in his employment. When the person is particularly able and the idea particularly worthy, such a match can lead to outstanding returns. We develop a stylized model that formalizes this idea. 3.1 Set-up We assume that each individual i is endowed with some innate ability a i. This ability determines the individual s earnings as a worker, which is normalized to equal a i. An individual also receives a business idea, the value of which depends on its intrinsic 6

9 merit and the fit of the idea with the individual s skills. Let denote the value of individual i s entrepreneurial idea, which is assumed to be independent of a i. If individual i becomes an entrepreneur, his earnings depend on the quality of his idea interacted with his ability:. (1) Paid wages are assumed to depend linearly on ability while entrepreneurial earnings are specified as its convex function to accommodate the Schumpeterian notion that highly talented individuals can generate great returns in entrepreneurship. Murphy, Shleifer and Vishny [1991], for example, argue that more talented individuals will choose to work where returns to talent diminish least and identify entrepreneurship as an activity that exhibits increasing returns to ability, in that having marginally greater talent leads to a significantly higher payoff. In their framework, this payoff is eventually limited because of diminishing returns in the baseline production function, while in our formulation it is the interaction between the individual s talent and the quality of the business idea that determines the returns from entrepreneurship We combine this theory of increasing returns to ability in entrepreneurship with some insights from the noisy learning model pioneered by Jovanovic [1979, 1982]. The learning process that we consider is not learning about one s productivity in a specific activity gained from experience in that activity, but learning from general work (and managerial) experience about how to assess the value of business ideas. We will see that the consequences of this kind of learning are similar to those examined by Jovanovic. More specifically, we assume that as individuals gain experience as workers, they become better able to assess the value of entrepreneurial ideas. They get to observe the returns from ideas developed in their firm, which provides useful information to project the value of ideas that subsequently come their way. We model this process in the simplest possible three-period framework, which corresponds to when individuals are young, middle aged, and old. For simplicity, we assume that an individual can be only a worker or an entrepreneur in each period. Workers receive a new potential entrepreneurial idea of quality in each period. A worker in period 1 only knows that the quality of his idea is a draw from a normal 7

10 distribution with mean 0 and variance. A worker in period 2 already has some experience, so together with a new idea drawn from the same prior distribution he receives a noisy signal about the quality of the idea:, (2) where and is independent of and a i. This signal is used to update the subjective distribution for according to the well-known Bayesian learning mechanism. Hence, the posterior distribution for is distributed normally with mean and variance, where (3). (4) Finally, workers in period 3 know exactly the quality of the (new) idea that they receive. We also assume that period 2 entrepreneurs learn the true quality of their ideas by the start of period 3 and may choose to return to paid work. In reality, feedback about the value of entrepreneurial ideas typically comes in the form of noisy realizations, such as earned profits. Consequently, entrepreneurs never truly learn the value of their ideas but become better informed about them over time as feedback accumulates. With the exception of Propositions 6 and 7 below this plays no role in our analyses, so we abstract from this aspect of reality in the model. When we link the theory to the data where we have bi-annual observations (see Section 5 below), we acknowledge that even older workers do not have perfect knowledge about their ideas and that learning about is done mostly in the first few years of tenure as entrepreneurs. 3.2 Becoming an Entrepreneur In our baseline model we want to abstract from other forces that may also be influencing the choice of entrepreneurship (such as risk preference and non-monetary compensation), so we assume that both workers and entrepreneurs are risk-neutral and maximize their expected earnings in each period. In period 1 the expected return from entrepreneurship equals 0 for all individuals, and so everyone begins their career as a worker. In period 2 a worker chooses to become an entrepreneur if his expected earnings from entrepreneurship exceed his wage from work: 8

11 , or. (5) In period 3 there is no uncertainty, hence the condition for becoming (or staying) an entrepreneur is or. (6) Since (and ) are independent of a i, it follows immediately from (5) and (6) that in every period more talented individuals are more likely to become entrepreneurs. Workers earn a wage equal to a i, hence it follows that: Proposition 1: The probability of becoming an entrepreneur is an increasing function of the worker s wage for workers in both period 2 and period 3. The rate of entrepreneurship and the ability of individuals who become entrepreneurs changes as a cohort ages. It can be shown (see Chernoff [1968, p. 227]) that is normally distributed with mean 0 and variance. Therefore, the probability of being an entrepreneur in period 2 can be expressed as:, (7) where is the cumulative distribution function of the standard normal distribution. In period 3 the quality of the business idea is known, so the probability of becoming an entrepreneur is given by. (8) Since, (7) and (8) imply that the probability of becoming an entrepreneur rises with labor market experience for workers at every ability level: Proposition 2: The probability of becoming an entrepreneur is an increasing function of the worker s labor market experience for any ability level (pre-entrepreneurship wage). While the probability of becoming an entrepreneur increases with age for all ability levels, this increase is proportionately greater for lower-ability workers. Formally, let denote the probability of becoming an entrepreneur in period 3 relative to period 2 for a given ability level a. Differentiating, 9

12 where, is the standard normal density function, and the inequality follows from the fact that is strictly decreasing in x for all x > 0. We have thus established: Proposition 3: The probability of becoming an entrepreneur in period 3 relative to period 2 is a strictly monotonically decreasing function of the ability level a. Proposition 3 implies that as workers gain experience, the probability of becoming an entrepreneur rises most on a percentage basis for the lowest wage earners. The intuition is as follows. Uncertainty about the quality of ideas in period 2 causes the distribution of posterior expected values of to be more concentrated around 0 (the mean of ) than the true values of. This compression has relatively little effect on the entry decisions of workers at high ability levels but has a much stronger effect at lower ability levels where the expected value of has to be high in order to justify entry. For example, consider two polar cases of individuals with (near) infinity and (near) zero ability. Among those with very high ability (close to infinity), nearly all who receive a positive signal will become entrepreneurs in period 2. Consequently, and. In contrast, almost nobody at very low ability (close to 0) will get a high enough signal in period 2 to justify becoming an entrepreneur. Hence, while tends to zero as a becomes very small, will tend to zero faster, and, which can be verified by applying l'hôpital's rule. Consequently, the lower the ability level then the greater the proportionate rise in the rate of entrepreneurship with age. 3.3 Earnings of New Entrepreneurs Since among those individuals who become entrepreneurs in period 3, we have by the well-known formula for the expectation of a truncated normal distribution. Thus, given ability a, the difference between the average earnings of new entrepreneurs and their prior wage is given by, (9) 10

13 where. The expression in the square brackets in (9) is decreasing in x (hence, increasing in a), so that the increase in average earnings from entrepreneurship will be an increasing function of the pre-entrepreneurship wage. In appendix we show that an analogous result also holds for individuals that become entrepreneurs in period 2. Therefore: Proposition 4: The difference between the average earnings of new entrepreneurs and their pre-entrepreneurship wage is positive and increasing in the pre-entrepreneurship wage. To understand the intuition for Proposition 4, note that the expected value of ideas for any non-zero ability level a is finite, so that as a approaches then the gain in earnings on the right-hand-side of (9) also approaches. Alternatively, as ability approaches 0, the gain in earnings on the right-hand-side of (9) approaches 0, as can be verified by applying l'hôpital's rule. Hence, higher-wage workers gain relatively more from becoming entrepreneurs than lower-wage workers. Since the probability of becoming an entrepreneur rises with a worker s wage (Proposition 1), it follows that Corollary 1: The average earnings of new entrepreneurs exceed the average earnings of workers. 3.4 Persistence in Entrepreneurship Workers who become entrepreneurs in period 3 know the quality of their entrepreneurial idea and hence will never exit from entrepreneurship. Workers who become entrepreneurs in period 2, on the other hand, do so on the basis of an imperfect signal about. Some fraction of them will therefore discover that the true quality of their idea was below the cutoff given by (6) and will return to paid work. Therefore, our theory implies: Proposition 5: The probability of exit from entrepreneurship is a decreasing function of an individual s labor market experience before becoming an entrepreneur. Proposition 5 relies on entrepreneurs in period 2 being uncertain about the quality of their ideas when they become entrepreneurs. As we noted earlier, over time entrepreneurs get (noisy) feedback that reduces their uncertainty about the value of their ideas. Consequently, pre-entrepreneurship labor market experience will have less effect 11

14 on the probability of exit among more experienced entrepreneurs, which we state as a corollary to Proposition 5: Corollary 2: The effect of pre-entrepreneurship labor market experience on the probability of exit from entrepreneurship will be a decreasing function of tenure in entrepreneurship. The same selection forces will affect the earnings of individuals that become entrepreneurs in period 2 relative to those that become entrepreneurs in period 3. Among those that become entrepreneurs in period 2, some will do so on the basis of signals that are greater than the true value of their ideas. If the true value of their ideas is such that, they will not be entrepreneurs in period 3. In the appendix we show that when we look only at period 2 entrants into entrepreneurship who remain entrepreneurs in period 3, for any given ability level the average value of their ideas will be higher than the average value of the ideas of period 3 entrants. Thus, for entrepreneurs of a given ability who stay in entrepreneurship long enough to learn about the true quality of their ideas, entrepreneurial earnings will be higher for those who became entrepreneurs at a younger age. We formulate this as: Proposition 6: For entrepreneurs of any ability level a, entrepreneurial earnings among entrepreneurs with sufficiently long tenure are a decreasing function of the individual s labor market experience before becoming an entrepreneur. Proof: See appendix. The same reasoning that underlies Corollary 2 applies as well to Proposition 6, except it generates the opposite prediction. Proposition 6 will hold most strongly for entrepreneurs with greater experience in entrepreneurship, as they are more selected in terms of their earnings. Consequently: Corollary 3: The negative effect of pre-entrepreneurship labor market experience on the earnings of entrepreneurs will be increasing (in absolute value) as a function of tenure in entrepreneurship. A robust finding in the labor literature is that labor earnings increase with labor market experience, even controlling for tenure on the job. This holds for workers in our sample as well. Proposition 6 implies the opposite result for entrepreneurs, especially 12

15 more experienced entrepreneurs, and thus provides an especially discriminating test of our theory. 4 Data and empirical methodology 4.1 The data The data come from the restricted-use Scientists and Engineers Statistical Data System (SESTAT) for the years 1995, 1997, and 1999 ( The National Science Foundation administered national surveys of individuals with (at least) a. U.S. bachelor s degree in science or engineering to gather employment, education, and demographic information. The surveys distinguish between workers and the self employed. The latter include individuals who answered in the affirmative the question whether their principal employment was in their own (incorporated or not) business, professional practice, or farm. In the empirical analysis, we exclude retired, unemployed, part-time workers, workers over age 65, and those who report salaries of zero as many of these individuals are likely to become self employed for defensive or life-style reasons. We also exclude occupations with almost no self employed, such as teaching, and occupations where sole proprietorships and partnerships are a common way of organizing activity, which includes health-related occupations, lawyers, judges, and agricultural occupations. 4.2 Identifying potential Schumpeterian entrepreneurs As noted in the introduction, two important general characteristics distinguish Schumpeterian entrepreneurs as agents of change from the general population of the self employed. First, Schumpeterian entrepreneurs tend to be more formally educated and disproportionately start their businesses in fast-growing and technologically sophisticated industries. Second, they tend to be exceptionally talented and also exploit ideas that they developed in their employment. The SESTAT dataset presents a better opportunity to focus on potential Schumpeterian entrepreneurs than general population surveys because it contains a nationally representative sample of highly educated scientists and engineers. But this paring of the population still may not be enough as even among highly educated 13

16 scientists and engineers there will be many who choose self employment for life-style and other reasons rather than novel ideas. To isolate the best candidates for Schumpeterian entrepreneurship, we pare down the sample to those that worked in and founded businesses related to their education. Respondents were asked whether their job was closely related, somewhat related, or not related at all to their educational degree, with about half of the individuals choosing the first answer. We restrict the sample to only those whose self-reported relation between the job (in paid work or self employment) and the highest educational degree was close and we refer to their employment as technology related. We exploit the longitudinal nature of SESTAT by restricting the sample to individuals that were surveyed at least twice over the three survey years of 1995, 1997, and To be included in our sample, individuals had to classify their job or business as closely related to their education in all years they were surveyed. This strategy has a number of advantages in trying to isolate the best candidates for Schumpeterian entrepreneurship. We do not have detailed industry breakdowns in the SESTAT data. However, a close relation between the job and education obtained in a specific science or engineering field is a proxy for working in one of the technologically sophisticated, high-growth industries from which successful entrepreneurial start-ups tend to come. Furthermore, a close relationship to a particular science and engineering field allows us to anchor both the job in paid employment and the job in entrepreneurship to a similar set of underlying technologies (and ideas), thus serving as a proxy for how much ideas derived from employment are exploited in entrepreneurship. Finally, Schumpeterian entrepreneurship is a story about superior talent. For scientists and engineers, there seems to be a strong correlation between ability and the closeness of the relationship between the job and formal training received at school. 5 Empirical results 5.1 Who Becomes An Entrepreneur: Testing Propositions 1, 2, and 3 Proposition 1 predicts that at any given level of labor market experience, higherearning workers will be more likely to become entrepreneurs. Furthermore, Propositions 2 and 3 predict that the fraction of workers at every earnings level that become 14

17 entrepreneurs will rise with labor market experience, but more so at lower earnings levels. We also expected that workers with longer tenure on the job would be less likely to change their jobs, including switching to self employment. We test these predictions by estimating the following probit equation:, (10) where is a 1-0 dummy variable equal to 1 if a worker in period t is self employed in period t+1, w t denotes paid earnings in period t, lex t denotes labor market experience prior to period t, 1 ten t denotes tenure on the job in period t, and X is a vector of demographic and occupational controls, including dummies for an MA degree, Ph.D. degree, professional degree, gender, white, married, handicapped, year, and 34 separate occupations (listed in Table A1 in the appendix). We also estimated (10) under the assumptions of the linear probability model and logistic distribution and the results were very similar. 2 The estimates of equation (10) are reported in Table 3. The estimates of β 1, β 2, and β 3 all conform with the predictions of the theory. The estimate of β 1 is positive and significant at the.05 level, indicating that the probability of becoming an entrepreneur rises with pre-entrepreneurial earnings. The estimate of β 2 is also positive and significant at the.01 level, indicating that the probability of becoming an entrepreneur rises with labor market experience. The estimate of β 3 is negative and significant at the.05 level, consistent with the probability of becoming an entrepreneur rising most with labor market experience for those at lower earnings/ability levels. The estimate of β 4 is negative and significant at the.01 level, consistent with greater tenure on the job lowering the probability of changing one s position. The demographic and occupational controls are jointly significant at the.01 level. The annual predicted probability of becoming an 1 This is measured as age in period t minus 22 for those with a college degree, age minus 24 for those with a masters degree, and age minus 27 for those with a Ph.D. 2 In all the analyses we experimented with various functional forms for the key variables, including quadratic and log specifications. With the exception of one analysis, the quadratic specification did not significantly improve the fit, and in that case a log specification best fit the data. We report only the bestfitting specifications. 15

18 entrepreneur ranges from 1.33 percent for workers under age 29 in the lowest labor market experience decile to 2.8 percent for workers above age 53 in the highest labor market experience decile. Table 3: Testing Propositions 1, 2, and 3 Marginal effects Independent variable at the mean Coefficient ** Log real salary St. Error Coefficient *** Labor experience St. Error Log real salary interacted Coefficient ** with labor experience St. Error Coefficient *** Tenure St. Error Coefficient *** Constant St. Error Other controls included Yes Number of observations (individuals) 26,232 (19,229) Log (pseudo)likelihood The dependent variable is a 1-0 dummy equal to 1 for individuals that moved from work in period t to self employment in period t+1. Other controls are master, Ph.D. and professional degree dummies, white, male, married and handicapped dummies, year dummies and 34 occupational dummies. Robust clustered standard errors are reported. *** indicates that the coefficient is significant at 1 percent level, ** at 5 percent level, and * at 10 percent level. Source: authors estimates using the NSF data. Proposition 1 predicts that the probability of becoming an entrepreneur should be an increasing function of earnings for every level of labor market experience. This requires β 1 + β 3 lex t to be positive for all values of lex t. The estimates of β 1 and β 3 are such that this condition holds for lex t up to 24 years. Since this corresponds to age 46 for those that entered the labor force at age 22, clearly the probability of becoming an entrepreneur is not increasing in earnings for older individuals. Proposition 2 predicts that the probability of becoming an entrepreneur should rise with labor market experience at every wage. This requires β 2 + β 3 w t to be positive for all values of w t. Based on the coefficient estimates, this condition is satisfied for all workers in the dataset. To better understand the relationship between the probability of becoming an entrepreneur and earnings, we divided observations into six groups according to whether the worker was above or below age 45 and whether the worker was in the bottom, 16

19 middle, or upper third of the earnings distribution (given his age). Table 4 reports the annual fraction of workers in each group that became entrepreneurs. Table 4: Fraction of workers becoming self employed by age and earnings Paid earnings Under 45 (A) 45 and over (B) Ratio: A/B Lower third Middle third Upper third Source: authors estimates using the NSF data For workers below age 45, the fraction of workers becoming an entrepreneur is highest among the highest paid workers, consistent with our theory, but is unexpectedly greater among those in the bottom than middle third of the wage distribution. For workers above age 45, the fraction becoming an entrepreneur in each earnings group is greater than for the respective earnings group below age 45, also consistent with the theory. But the fraction of workers becoming entrepreneurs among the older workers is unexpectedly highest in the lowest third of the earnings distribution. Note, though, that the fraction of workers becoming entrepreneurs increases with age more than proportionately at lower ability levels, consistent with Proposition 3. Finer breakdown of the age categories reveals that at the youngest ages entry rates rise monotonically with paid wages as predicted by our theory. However, as workers age the increase in the entry rate at the lowest wage level is so sharp that entry rates in this group unexpectedly exceed those even for the highest wage group at the oldest ages. We will return in the next section to address why we think this departure from the model occurs. 5.2 Earnings of New Entrepreneurs: Testing Proposition 4 Proposition 4 predicts that workers that become entrepreneurs will increase their earnings, with the increase greater for workers with greater pre-entrepreneurship earnings. This suggests the following empirical specification:, (11) where se t+1 are the earnings of workers in period t that are self employed in period t+1, w t are the earnings of the workers in period t, and ε t is the disturbance term. If earnings from work, w t, were purely a reflection of an individual s ability, then there would be no problem estimating equation (11). But earnings from work will vary 17

20 from year to year due to factors such as the availability of overtime work. Annual earnings are also likely to be measured with error. Both factors will contribute to a negative correlation between the dependent variable in (11) and earnings, which in turn will cause OLS estimates of β to be negatively biased. The same would be true if the dependent variable in (11) were ln w t+1 ln w t, which might be expected to be unrelated to ln w t if earnings were measured correctly and did not contain a transitory component. Symptomatic of the difficulties of estimating equation (11) is that when we did estimate equation (11) with ln w t+1 ln w t as the dependent variable, the estimate of β was negative and significant at the.01 level. 3 If the random component of wages due to transitory factors and measurement error are uncorrelated across years, we can estimate consistently the coefficients of (11) using ln w t-1 as an instrument for ln w t. We report these instrumental-variable estimates in Table 5. They are based on the 108 individuals in our dataset that were surveyed in all three years and reported being a worker in 1995 and 1997 and self employed in We also present in Table 5 the coefficient estimates of the same model restricted to the 6,912 individuals that reported being a worker in all three survey years, with ln w t+1 substituted for ln se t+1. To increase the precision of our estimates, we drop all the demographic and occupational control variables, which perhaps not surprisingly are jointly insignificant. The estimate of coefficient β for the regression with ln se t+1 ln w t as the dependent variable is positive and significant at the.01 level, indicating that the increase in earnings from entrepreneurship is greater for workers with higher (preentrepreneurship) earnings, consistent with the theory. The estimates imply that workers at the median of the paid earnings distribution gain about 3.3 percent in earnings from the transition to entrepreneurship, while workers at the 75 th percentile gain about 14.2 percent. In contrast, workers at the 25 th percentile lose about 8 percent of their earnings after the transition. As we will discuss, this may be due to the same factors that we conjecture are behind the high rate of entrepreneurship among older individuals at low 3 Estimating equation (11) with ln se t+1 ln w t as the dependent variable leads to the estimate of β that is still negative, but half of the magnitude of the estimate for workers and statistically not significant. 18

21 earnings levels. Note that when ln w t+1 ln w t is used as the dependent variable and the model is estimated using the same instrumental-variable strategy, the estimate of β is negative and significant at the.01 level. This suggests that the estimates for those becoming entrepreneurs are not an artifact of the estimation procedure or form of the model. Table 5: Testing Proposition 4 First stage regression Self employed in t+1 Paid workers in t+1 Paid workers in t-1 and in t Log earnings in t Coefficient *** *** Log earnings in t-1 St. Error Coefficient *** *** Constant St. Error R squared Second stage regression Log earnings in t+1 - Log earnings in t Log earnings in t Coefficient *** *** (instrumented) St. Error Coefficient *** *** Constant St. Error Number of observations 108 6,912 Joint significance test statistic (Prob > F) Two-stage least squares estimation is employed, with log earnings in period t estimated in the first stage as a function of log earnings in period t 1 and labor experience. *** indicates that the coefficient is significant at 1 percent level, ** at 5 percent level, and * at 10 percent level. Source: authors estimates using the NSF data. The corollary to Proposition 4 predicts that the average earnings of new entrepreneurs will exceed the average earnings of workers. We can test this prediction without having to resort to the instrumental-variable approach by estimating the following equation:, (12) where ln earn t+1 is the log of the individual s wage or self-employment earnings in period t+1, D t+1 is a 1-0 dummy variable equal to 1 for individuals that shift from work to self employment in period t+1, lex t is labor market experience in period t, X is the set of demographic and occupational controls, and ε t is the disturbance term. The variable lex t and the demographic and occupational controls are included to allow for general factors expected to influence earnings for all individuals. Propositions 3 and 4 suggest that the difference between the earnings of new entrepreneurs and workers may be declining in 19

22 labor market experience because proportionately more low-ability workers enter entrepreneurship in period 3 than period 2 and the gains from entrepreneurship increase with ability. We included the interaction term between D t+1 and lex t in regression (12) to allow for this possibility, which would be reflected in β 2 less than 0. The estimates of this model are reported in Table 6. Considering first the controls, the estimate of the coefficient of labor market experience, β 3, is positive and significant at the.01 level, as would be expected. The demographic and occupational controls are jointly significant at the.01 level, also as expected. In terms of the theory, the estimate of β 1 is positive and significant at the.01 level, as predicted. It implies that among individuals with no prior labor market experience, the newly self employed earn about 20 percent more than comparable workers. The estimate of β 2 is negative and significant at the.01 level and implies that each prior year of labor market experience decreases the self employed-worker earnings differential by about 0.8 percent. Consequently, for those with 25 years of labor market experience, the newly self employed earn about the same amount as comparable workers, and for yet older individuals the newly self employed earn less than workers. Table 6: Average earnings of new entrepreneurs versus workers Dependent variable Log t+1 earnings Coefficient *** Mobility dummy St. Error Mobility dummy interacted Coefficient *** with labor experience St. Error Coefficient *** Labor experience St. Error Coefficient *** Constant St. Error Other controls included Yes Number of observations (individuals) 26,270 (19,250) R-squared Mobility dummy equals 1 if the individual moved from paid work in period t into self employment in period t+1. Other controls included are master, Ph.D. and professional degree dummies, white, male, married and handicapped dummies, year dummies and 34 occupational dummies. Reported are robust clustered standard errors. *** indicates that the coefficient is significant at 1 percent level, ** at 5 percent level, and * at 10 percent level. Source: authors estimates using the NSF data 20

23 Our theory predicts that new entrepreneurs earn more than comparable workers, albeit the differential may be less for those with greater labor market experience. The lower relative earnings of the newly self employed at older ages could reflect a downward bias in the measure of the earnings of the self employed relative to workers. In the next section we provide an alternative explanation for this departure from the theory, attributing it to the same factors we conjecture are behind the high rate of entrepreneurship among older individuals at low earnings levels. Elfenbein et al. [2008] find that in the SESTAT dataset, employees of smaller firms are significantly more likely to enter self employment than employees of larger firms. This also holds in our restricted, technology-related sample. Workers in small firms also earn less than workers in large firms. Consequently, if we include firm size dummies in equation (12) to control for the type of workers more likely to become self employed, the estimates conform even more with our theory. The new estimates of β 1 and β 2 together imply that the newly self employed now earn more than comparable workers among individuals with up to 30 years of labor market experience. 5.3 Persistence in Entrepreneurship: Testing Propositions 5 and 6. Proposition 5 predicts that the greater the amount of labor experience before becoming an entrepreneur, the less likely an entrepreneur will return to being a worker. For continuing entrepreneurs in our sample we do not know exactly when they became self employed, but we can subtract tenure in the current business from labor market experience to obtain a proxy for labor market experience prior to entrepreneurship. Corollary 2 implies that tenure in entrepreneurship will also condition the effect of preentrepreneurship labor market experience on the probability of exiting entrepreneurship. We test these predictions by estimating the following probit model: (13) where is a 1-0 dummy equal to 1 if a self-employed individual in period t is a worker in period t+1, log_prior_lex t is the log of labor market experience in period t prior to the current business, log_ten t is log entrepreneurial tenure in period t, log_prior_lex t * log_ten t is the corresponding interaction term, and X is the set of demographic and 21

24 occupational controls. We expect β 1 to be negative and β 3 to be positive. Studies have found that longer tenure in entrepreneurship also reduces exit rates and thus we expect β 2 to be negative. 4 The estimates of (13) are reported in Table 7. Consistent with our predictions, the estimates of β 1 and β 2 are negative and significant at.01 level while the estimate of β 3 is positive and also significant at the.01 level. The demographic and occupational controls are jointly significant at the.01 level. The estimate of β 1 implies that among new entrepreneurs (with zero log tenure), the probability of exit declines by about 1.4 percent for each 10 percent rise in pre-entrepreneurship labor experience. 5 The estimate of β 3 implies that half of this effect is dissipated by the time entrepreneurs have accumulated five years of tenure in entrepreneurship. Table 7: Testing Proposition 5 Independent variable Log labor experience, minus tenure Coefficient *** (pre-entry experience) St.Error Coefficient *** Log tenure in business St.Error Coefficient *** Interaction term St.Error Coefficient *** Constant St.Error Number of observations (individuals) 1,510 (1,181) Log pseudo-likelihood The dependent variable is the dummy which equals 1 if the individual moved from self employment in period t to work in period t+1. Other controls include master, Ph.D. and professional degree dummies, white, male, married and handicapped dummies, year dummies and 34 occupational dummies. Robust clustered standard errors are reported. *** indicates that the coefficient is significant at 1 percent level. Source: authors estimates using the NSF data. 4 Both labor market experience and tenure were expected to have diminishing effects on the probability of exit, which motivated the log specification. We also estimated (13) using a quadratic specification for both variables (and their interaction) and the results were similar. 5 We also estimated equation (13) for just new entrepreneurs (without the controls for tenure). Consistent with the estimates for the full sample, the estimated effect of pre-entrepreneurship labor market experience on the probability of exit was negative and significant. Moreover, it was of a similar magnitude (about 1.6 percent decline in the exit probability for each 10 percent rise in labor market experience) to the estimate for the full sample. 22