Self-Employment Dynamics and

the Returns to Entrepreneurship∗

Eleanor W. Dillon Christopher T. Stanton

Arizona State University Harvard Business School,

NBER, and CEPR

February 12, 2016

Preliminary

The Latest Version Can Be Found Here

Abstract

In a cross-section of workers the median individual in self-employment earns less than themedian paid worker, but the highest-paid entrepreneurs earn more than employees. Smallbusiness owners and others in self-employment have the option to transition back to paid work.If they are initially uncertain about their entrepreneurial abilities, this option increases the life-time expected value of entering self-employment relative to the expected pay in a single year.This paper first documents that moves between paid work and self-employment are commonand consistent with experimentation to learn about entrepreneurial ability. This motivatesestimating the expected returns to entering self-employment within a dynamic lifecycle modelallowing for non-random selection in and out of self-employment and gradual learning aboutentrepreneurial ability. The option value of returning to paid work is found to be a substan-tial part of the value of entering self-employment. Accounting for this option decreases theinferred non-pecuniary benefits of entrepreneurship necessary to reconcile observed earningsand workers’ labor choices.

∗ewdillon@asu.edu, christopher.t.stanton@gmail.com. We thank Kelly Bishop, Luis Garicano, Bill Kerr, Ed-ward Lazear, David de Meza, Robert Miller, Ramana Nanda, Luis Rayo, Yona Rubinstein, Nathan Seegert, GalinaVereschagina, and luncheon participants at the London School of Economics for helpful comments. All errors areour own.

1 Introduction

The median self-employed worker earns less than the median wage worker.1 Several candidate

explanations have been advanced to explain this gap, including non-pecuniary benefits from being

one’s own boss, mis-reported self-employment earnings, and improper accounting for nonrandom

selection. Given that many spells in self-employment last only a few years, dynamic selection is

likely of particular importance. However, most existing analyses of the self-employment decision

calculate static (cross sectional) returns: individuals are either self-employed or paid workers. With

the exception of Manso (2014), there has been little work that documents differences in the value

of paid work and self-employment after accounting for the ability to change sectors.2

One possible motivation for movements between paid work and self-employment is uncertainty

about sector-specific ability. Workers may have an incentive to experiment with self-employment

and learn more about their skills, knowing that their overall downside risk is limited because they

have the option to return to paid work. Several stylized facts from the Panel Study of Income

Dynamics (PSID) suggest that workers are initially uncertain about their entrepreneurial ability,

but can learn about it by working in self-employment.3 First, many workers in the PSID have spells

of self-employment, but few stick it out and remain self-employed long-term. Nearly a quarter of

the individuals in the PSID enter entrepreneurship at some point in their careers, but almost

half of those who do return to the paid sector within five years. Workers also seem unable to

forecast whether they will be good entrepreneurs: the mean person who moves into entrepreneurship

experiences a drop in earnings relative to his last paid job. Third, the least successful entrepreneurs

are more likely to quickly return to paid work, while those with higher earnings are more likely to

persist in entrepreneurship.

1This gap is documented in the Panel Study of Income Dynamics data used here. Studies using a variety of datasources also show that the median individual in self-employment earns less than the median wage worker. See, forexample, Evans and Leighton (1989), Hamilton (2000), and Levine and Rubinstein (2013). Work on the valuation ofprivate businesses shows that returns are no higher than investments in publicly traded equities, despite their higherrisk (Moskowitz and Vissing-Jorgensen (2002), Hall and Woodward (2010)).

2Using a matching on observables approach, Manso (2014) shows that individuals who enter self-employment atsome point, including those who eventually return to paid work have higher total lifetime earnings than those whonever work for themselves.

3In this paper we define entrepreneurs as anyone who reports working for themselves in their main job; en-trepreneurship and self-employment are therefore used as synonyms. We discuss this choice in the data section.

1

If workers experiment in this way, then simple comparisons of earnings across sectors do not

provide unbiased estimates of the population average return to entering entrepreneurship. Suppose

that one calculates the expected value of entering entrepreneurship by comparing earnings for

current entrepreneurs and current paid workers, controlling for observed worker characteristics. This

cross-sectional estimate may overstate or understate the expected value of entering entrepreneurship

because of two competing forces. On the one hand, a cross-section of self-employed workers includes

some new entrepreneurs with low earnings who will eventually return to paid work. For these

workers, their current low earnings understate their expected remaining lifetime earnings. On the

other hand, successful entrepreneurs are over-represented in a cross-section of self-employed workers.

Survivorship differences may inflate cross-section estimates relative to what an average person can

expect to gain from entering entrepreneurship.

In this paper we assess the relative importance of these two sources of bias and quantify the

option value of experimentation in entrepreneurship. To do so, we specify a simple model of

earnings and labor sector choice that captures the key incentives driving movements in and out

of self-employment over the lifecycle and estimate the it using data from the PSID. In the model,

risk-neutral individuals have heterogeneous, sector-specific abilities. Workers do not know their

entrepreneurial ability with certainty and must learn about it by working in self-employment. We

allow workers to have positive or negative tastes for working in entrepreneurship. We identify these

utility parameters using data on moves between paid work and self-employment, along with data

on earnings.

We use these estimates to measure the present utility value of choosing to work in each sector

as a function of the expected discounted stream of flow earnings, any non-pecuniary benefits or

costs from entrepreneurship, and entry costs into self-employment. We can then project expected

lifetime earnings for each worker in the PSID, conditional on sector choice next year and allowing

workers to move between sectors in future years to maximize their utility. The average difference

between the expected discounted lifetime earnings conditional on choosing entrepreneurship next

year and the earnings from choosing paid work is an unbiased estimate of the monetary returns

to entrepreneurship. We also calculate these projections under the counterfactual assumption that

2

workers must remain in one sector in all future periods. The difference between the estimated

and counterfactual value of entrepreneurship defines the value of the opportunity to resume paid

employment.

In addition to documenting several new stylized facts about transitions in and out of self-

employment over the lifecycle, we reach three key conclusions about the returns to entrepreneurship.

First, selection bias in cross-sectional comparisons is important, particularly at the mean. In the

PSID, the mean earnings for a 32 year old who is observed in entrepreneurship is $75,000 per year,

26% more than the mean earnings for workers observed in paid jobs.4 These figures reflect workers’

strategic selection into each sector. In contrast, we estimate that if all 32 year old workers chose

to work as entrepreneurs, their mean earnings would be $59,000, virtually the same as the mean

earnings of $58,500 that we project these 32 would earn if they all chose the paid sector. On average,

32 year old workers could expect to earn 0.1% more working as entrepreneurs for a year than they

would in a paid job.

However, when we consider expected earnings in all future years, the option value of learning

entrepreneurial ability makes self-employment more attractive. For 32 year old workers, static

lifetime earnings, assuming that workers must remain in the same sector in all periods, are 1.5%

higher at the mean and 0.6% higher at the median in entrepreneurship relative to paid work.

However, expected lifetime earnings from the dynamic model are 11.2% higher at the mean and

12.2% higher at the median for workers who enter entrepreneurship relative to those who remain

in the paid sector. This difference reflects the long right tail of entrepreneurial earnings. While

most workers earn the same or somewhat less in entrepreneurship as they would in paid work, some

workers earn far more. The monetary value of learning whether you are one of those successful

entrepreneurs outweighs the monetary opportunity cost of spending several years outside of the

paid sector. On average, the option value of experimenting and then selecting back into paid work

accounts for 11% of the expected lifetime value of entering entrepreneurship. This share differs over

individuals and over the lifecycle.

Finally, tastes for self-employment vary considerably across workers. Modeling this heterogene-

4The modal entrepreneur enters self-employment in his early 30s, making this a particularly relevant comparisonage; the model allows discussion of these statistics at other ages as well.

3

ity is important for matching observed patterns of entrepreneurial choice. After accounting for

the dynamic option value of experimenting in entrepreneurship, and in contrast to most of of the

previous literature, we estimate that the mean utility from working in entrepreneurship is negative.

The average worker would give up $10,000 per year to avoid working as an entrepreneur. However,

27% of the population has a positive non-pecuniary benefit from entrepreneurship.

These findings build on earlier work to improve our understanding of workers’ choices about

entrepreneurship. Earlier studies (Evans and Leighton (1989), Hamilton (2000), Moskowitz and

Vissing-Jorgensen (2002), Hall and Woodward (2010)) have attempted to explain why the returns

to entrepreneurship might be small or negative. After accounting for dynamic considerations, entry

into entrepreneurship is no longer a puzzle.5 In fact, observed sector choices reverse findings about

the overall importance of non-pecuniary considerations or underreporting for tax reasons.6

The next section provides a motivating example to illustrate the importance of the option to

switch sectors. Section 3 describes the data and presents stylized facts that suggest the importance

of considering returns to entrepreneurship in a dynamic context. Section 4 presents the full model

while section 5 presents the estimation algorithm. Section 6 presents our estimates and discusses

fit, section 7 evaluates the value of entering entrepreneurship, and section 8 concludes.

2 Motivation

To illustrate why the option to change back to paid work affects inference about earnings in self-

employment, consider a simple problem of choosing entrepreneurship for a risk neutral individual

who lives 1 period. His earnings in entrepreneurship equal his entrepreneurial ability, ηi, which has

an ex-ante normal distribution with mean µi and standard deviation σ2. If he works in the paid

5 Previous theoretical contributions, especially Vereshchagina and Hopenhayn (2009), discuss the role of en-trepreneurial selection in a dynamic model. In related work, Manso (2014) articulates that cross-sectional estimatesof entrepreneurial earnings are biased because of dynamic concerns. He then shows that realized lifetime earningsin the NLSY79 are higher for workers who spend at least some time as entrepreneurs, consistent with the findingshere. Relative to Manso’s work, this paper has the advantage of identifying the option value of experimenting inentrepreneurship separately from bias due to selection.

6The paper also fits into the large literature on who selects into entrepreneurship and the smaller literature on theexperiences of former entrepreneurs. Hamilton (2000), Evans and Leighton (1989), and Bruce and Schuetze (2004)provide some evidence on the return to self-employment experience in wage work. Caliendo and Uhlendorff (2008),Caliendo et al. (2010), Levine and Rubinstein (2013), and Hurst and Pugsley (2011) discuss the heterogeneity ofself-employed workers and the reasons they select into self-employment.

4

sector he earns $1200 with certainty. Without non-measured benefits from entrepreneurship, in a

single period model, the worker chooses entrepreneurship if and only if µi, his expected earnings,

exceed $1200.

Now consider a two period model with the option to move between sectors. Suppose en-

trepreneurial ability is unknown before entering entrepreneurship, but learned with certainty after

one period. In this case, the threshold mean required to choose entrepreneurship is weakly less than

$1,200. Expected lifetime earnings from choosing entrepreneurship in the first period equals

ηi + E {max (ηi, $1, 200)} .

The second term captures the option to switch from entrepreneurship back into the paid sector

if earnings in entrepreneurship do not justify continuation. For any possible value of of σ2, it is

possible to solve for the critical value µ∗i such that µ∗i +E {max (ηi, $1200)} ≥ $2400.7 When σ = 0,

the critical value µ∗i = 1200, the certain earnings in the paid sector. As σ increases, the mean of

the distribution required to experiment in entrepreneurship declines monotonically, reflecting the

increased option value associated with variability.

This exercise also illustrates career dynamics across paid and self-employment. Individuals who

are revealed to have ηi ≤ $1200 return to the paid sector while those who have ηi > $1200 remain

entrepreneurs. The full model makes this structure more realistic, allowing for non-pecuniary ben-

efits (or costs) from self-employment, costs of entering entrepreneurship, sector-specific returns to

experience, and unobserved heterogeneity. Before getting there, the next section shows that career

dynamics observed in the data are consistent with this declining hazard of leaving entrepreneurship

over time.

7Using the moments of the truncated normal distribution, this is µ∗i +{µ∗i + σ

φ(c)1−Φ(c)

}(1−Φ (c))+ 1200Φ (c) =

$2400 where c = (1200− µi) /σ.

5

3 Data and Summary Statistics

3.1 The Sample

The data are from the 1976-2011 waves of the Panel Study of Income Dynamics (PSID). The PSID

interviewed respondents annually until 1997 and bi-annually from 1999 onward. The long panel

structure of the PSID allows us to observe workers before, during, and sometimes after spells of

entrepreneurship. The original PSID sample includes a representative group of American households

in 1968 and an oversample of low-income households. The PSID has continued to interview members

of these households, their offspring, and individuals who marry into these families. Over time, some

original families have been dropped and the sample has been augmented with several samples of

recent immigrants to better reflect the current mix of US households. The baseline sample includes

all households interviewed in each year; PSID-constructed weights are used to adjust for probability

of inclusion in the survey. These changes in the sample, along with workers aging into and out of

the workforce and occasional non-response, create an unbalanced panel.

From this baseline sample, the estimation sample is constrained by the need to keep track of ac-

cumulated work experience in each sector. Individuals are included in the sample only when we can

follow their work experience starting at age 25 or earlier.8 Because we estimate individual-specific

earnings effects, the most restrictive inclusion criteria further reduces the sample to individuals who

report at least 3 years of labor earnings in either sector.9 In this paper the sample is restricted

to men.10 Finally, to abstract from early retirement behavior and schooling decisions, we include

8Both survey responses and the panel dimension of the data are then used to construct measures of accumulatedexperience in each sector. In other words, we keep all workers who enter the sample by age 25, plus any worker whoenters the sample at an older age but reports being in their current job since before they were 25. The PSID onlybegan asking about job tenure in 1976, but observed paid or self-employed work from 1968-1975 help to constructexperience measures. In the first year a worker is observed, we initialize their sector experience using their reportof how long they have been at their current job. For example, if a worker enters the sample working in the paidsector and has been working at that job for 8 years, we say he has 8 years of paid experience and no entrepreneurialexperience. Experience in each sector is then updated with observed work going forward.

9The 3 year cutoff is used for the first stage of estimation when only short-panels are needed to recover variancecomponents through moment based estimation. The second stage of the structural model requires estimates ofindividual earnings abilities as a state variable. For this exercise, 5 years of data are required.

10Women in the PSID are far more likely to persist in self-employment despite low relative earnings, suggestinga larger role for non-pecuniary considerations or a fuzzy divide between self- and non-employment. Women’s en-trepreneurial decisions are a valuable area for future research, but it seems clear that men and women should not bepooled to estimate a single model.

6

only men between the ages of 22 and 55. After these restrictions, the sample includes just under

7,000 men. On average, we observe 10 years of earnings for each individual in the sample, with a

maximum of 29 observations spanning 36 years.

3.2 Identifying Entrepreneurs

We define an entrepreneur as someone who is self-employed in their main job.11 This definition

is used, rather than one based primarily on business ownership, because entrepreneurship here is

modeled as a labor supply choice that may also represent a financial investment. Light and Munk

(2015) investigate the difference between self-employment and business ownership in more detail.12

Table 1 describes workers who ever spend time in entrepreneurship and those who only ever

work in the paid sector. Of the set of individuals with experience in entrepreneurship, 82% own

a business at some point during the sample. Further, 78% of entrepreneurs own a business in

any given year (reported from calculations behind Figure 1). The rest work as contractors or are

otherwise self-employed without owning a business.13 Entrepreneurs are more likely to be white

and slightly better educated on average than the paid-only group, but otherwise look quite similar

to other workers.

Moves in and out of entrepreneurship are quite common. The average member of the sometime

entrepreneur sample is observed for 7 years in paid work and only 5 years in entrepreneurial work.

The median entrepreneur is 32 when he first enters entrepreneurship. Figure 1 plots the cumulative

transitions back to the paid sector and changes in business ownership by years since workers first

enter entrepreneurship. Overall 40% of workers who experiment with entrepreneurship return to

the paid sector within 5 years. Experienced entrepreneurs are substantially more likely to own a

11In most cases, this simply means the worker is self-employed. The PSID asks workers who hold more than onejob simultaneously to identify their main job, which tends to be the job that accounts for the majority of hoursworked and earnings. We classify individuals who work for someone else on their main job and for themselves in aside job as paid workers, and vice versa.

12In the later part of the PSID sample, business owners are asked if they work in their business. In these lateryears we observe that some business owners who report working in their business and having only one job nonethelessreport working for someone else. This combination of responses is somewhat puzzling, but on the whole we wouldlike to classify these workers as entrepreneurs throughout the sample. To do so consistently, we also identify workersas entrepreneurs when their first report of owning a business coincides with a job change. This rule captures mostof the respondents who report working in their businesses in the later half of the sample.

1321% of workers who always work in the paid sector also own businesses in at least one year, but these businessesare run on the side while working for someone else on the main job.

7

business and that business is more likely to be incorporated, which Levine and Rubinstein (2013)

find is a good proxy for business sophistication and profits.

However, using business ownership to define entrepreneurship likely understates experimentation

in self employment. As shown in Table 2, there is a composition change of business ownership

that occurs with experience in self employment both because workers who enter entrepreneurship

as business owners are more likely to persist and because some workers who begin without a

business establish one several years into self-employment. This ability to ”upgrade” into business

ownership is a likely consequence of experimenting in self employment without first paying the costs

of incorporation.

3.3 Earnings in Paid Work and Entrepreneurship

Calculating earnings for the self-employed requires several data cleaning steps. In the PSID, paid

workers are asked about their total labor earnings from wages, salaries, bonuses, commissions, and

tips. Most owners of incorporated businesses are also asked about total labor earnings, but owners

of unincorporated businesses are instead asked about their net profit from that business. Our

first choice for both paid and self-employed workers is to use reported annual labor earnings. For

entrepreneurs who do not report labor income but do report profit from a business, reported net

profit is used as their labor earnings.14

One potential drawback is that using net profits will overstate labor earnings in cases where

workers have also invested substantial financial capital in their businesses. Information about

business assets are too incomplete in the PSID to allow us to adjust net profits for a reasonable

return to capital. However, the data that are available suggest that these adjustments would be

small. As shown in Table 3, the median worker who opens a business upon entering entrepreneurship

invests $5,000 over the first 1-5 years. This number likely overstates the median entry investments

for all self-employed workers, since workers who become self-employed without reporting ownership

of a business are not asked about capital investments. Even at the 90th percentile, unincorporated

14Because both of these measures are for annual earnings, earnings for years when individuals spent part of theirtime in the paid sector and part in entrepreneurship are excluded from earnings regressions.

8

business owners invest only $59,000. At a 5% rate of return, that suggests self-employment earnings

may be overstated by only about $2,900 per year.

Both labor income and business profits are reported over the past calendar year, while job tenure

is reported as of the survey date.15 Because work hours are difficult to define for self-employed

workers, earnings are measured at the weekly level using annual earnings divided by reported weeks

worked.

Earnings in entrepreneurship are more variable than earnings in the paid sector. Figure 2

plots the distribution of real weekly earnings in each sector. The distribution of earnings for

workers currently in entrepreneurship is flatter than for paid workers, with more weight on the

lowest values and a thicker long right tail. Controlling for age, sex, race, and education, we find

that the mean of earnings in entrepreneurship is somewhat higher than the mean in wage work,

but the median is lower.16 Table 4 presents OLS and quantile regressions of real weekly earnings

including an indicator for being an entrepreneur and basic demographics. In the sample, the average

entrepreneur earns $403 per week more at the mean, but $162 less at the 25th percentile and $43

less at the median. These regressions predict mean weekly earnings of $1,075 and median earnings

of $720 for paid earners, so these gaps represent a 37% increase at the mean and a 6% drop at the

median, respectively.

Past literature uses estimates at the median to ask why some individuals stay in self-employment

even though they appear to earn less than they would earn in paid work. The benchmark non-

pecuniary benefit at the median in the static model necessary to rationalize this decision is thus

about 6% of paid earnings. It will be shown later that the average non-pecuniary benefit from the

full model deviates significantly from this figure.

15To estimate the effect of sector experience on earnings we construct experience as of the beginning of the lastcalendar year, using reports from both the current and previous waves of the survey.

16Hamilton (2000) finds a similar result in the Survey of Income and Program Participation, using a somewhatdifferent measure of earnings in entrepreneurship.

9

3.4 Learning and Selection In and Out of Entrepreneurship

Recall that Figure 1 showed that moves from entrepreneurship back to the paid sector are quite

common. Figure 3 illustrates that these transitions back to the paid sector are not random. The

top line of this figure plots the experience profile of earnings in entrepreneurship for workers who

remain in entrepreneurship for 6 or more years. This profile is strictly above the experience profile

of workers who leave within 5 years, which is in turn higher than the earnings in entrepreneurship

for workers who leave after one year. Workers who persist longer in entrepreneurship earn more

in that sector from their first year. The same pattern holds when we consider entrepreneurial

earnings relative to expected earnings in the paid sector. Figure A1 plots the same series of observed

earnings in entrepreneurship, now relative to projected counterfactual earnings had these individuals

instead worked in the paid sector. The estimates we use to construct these counterfactual earnings

are described in Section 6.1.17 All entrepreneurs earn less in their first year of entrepreneurship

than we would expect them to earn had they remained in paid work. Those who will remain in

entrepreneurship for at least 6 years begin earning more than their counterfactual paid earnings by

their second year of entrepreneurship, while individuals who will return to the paid sector within 5

years continue to earn less than their projected paid earnings.

This negative correlation between entrepreneurial earnings and the probability of leaving en-

trepreneurship could be consistent with a learning model or with a model in which workers know

their abilities with certainty but experience large sector-specific taste shocks. However, this expla-

nation is unlikely and a number of data points suggest learning is important. First, the majority

of workers in their first year of entrepreneurship earn less than they did in their last year of paid

work, suggesting the willingness to take on possible low earnings for the potential to learn about

future opportunities.18 New entrepreneurs earn $1,200 less per year at the mean and $6,000 less at

the median than their most recent year of paid earnings, despite benefiting from an additional year

17These counterfactual earnings are constructed using only parameters from the first stage of estimation, describedbelow. While these estimates are designed to be robust to the kinds of selection we hypothesize their identificationdoes not depend on these model assumptions. The figure is also similar when using residual earnings.

18These estimates are based on 190 individuals for whom we observe earnings in their first year of entrepreneurshipand paid earnings within the previous two years. These gaps are consistent with the differences between observedentrepreneurial earnings and projected paid earnings, which we can calculate for all new entrepreneurs, plotted inFigure A1.

10

of experience. The median new entrepreneur earns 17% less than he did in his last year of paid

work.

Second, if workers have some understanding of their entrepreneurial abilities even before entering

entrepreneurship, then we would also expect workers with high entrepreneurial abilities to be willing

to invest more when starting their businesses. The last three columns of Table 3 look at how

initial investments in business covary with individual average paid earnings and entrepreneurial

earnings. These averages for each sector are essentially individual fixed effects estimated from

realized earnings in each sector with an adjustment for sampling variation. Paid and entrepreneurial

abilities are positively correlated, so initial investments should be rising in paid earnings whether

or not individuals also have advanced knowledge of their entrepreneurial ability. The ratio of

entrepreneurial earnings to paid earnings is thus a better measure of outcomes in entrepreneurship.

The third column of Table 3 shows that the ratio of mean entrepreneurial earnings fixed effects to

mean paid earnings fixed effects rises slightly with investment amounts given that investment is

positive. Individuals who invest between $1 and $5,000 in their business have entrepreneurial ability

that is 2% lower than their paid ability on average while those investing more than $65,000 have

entrepreneurial ability that is 5% higher than their paid ability on average. However, as shown in

the last column, these differences in averages are dwarfed by the standard deviation of this ratio. In

addition, if larger initial investments increase the probability of entrepreneurial success or if there

is a return to capital investment that is being omitted, then initial investments will be positively

correlated with the entrepreneurial earnings fixed effect even if workers do not know this ability in

advance. Finally, individuals who make no investment have earnings in entrepreneurship that are

about 5% higher than paid earnings, which is about the same ratio as individuals who invest more

than $65,000.

Once workers have entered entrepreneurship and observed their earnings in that sector, those

with low relative entrepreneurial earnings are far more likely to return to the paid sector. We see

much weaker evidence that workers who have not yet worked as entrepreneurs have any advance

knowledge of their potential earnings in that sector.

With these stylized facts about the decision to become an entrepreneur, we turn now to a model

11

of behavior that allows us to measure preferences for money and entrepreneurship. The model will

be used to measure returns and the option value of experimentation.

4 Model

In each period t = 0, 1, ..., Ti, starting after the last year of schooling and continuing to retire-

ment, risk-neutral individual i chooses between supplying labor in the paid sector (dit = 0) or the

entrepreneurial sector (dit = 1). Once the sector is chosen, earnings shocks are realized and flow

payoffs are received. This assumption means that the individual makes sectoral choices without

knowing earnings in advance, and adverse shocks cannot be escaped before their realization. This

is consistent with a world in which there is some earnings risk over the course of a period, perhaps

from job loss, or where earnings, even in the paid sector, depend partially on bonuses or profit

sharing.

Our assumption of risk neutrality is reasonable if workers have some ability to smooth con-

sumption. The option to return to paid work limits individual downside risk from entering en-

trepreneurship to a few periods. With smoothing, low earnings in entrepreneurship today affect

total consumption only through the lifetime budget constraint. As a result, and consistent with the

data presented in Hurst et al. (2014), flow consumption changes are likely to be much smaller than

the change in flow earnings for those who have low earnings in entrepreneurship. Model complexity

and data issues make adding risk aversion difficult for this paper. Adding endogenous savings and

consumption decisions greatly complicates the model solution. The PSID includes very limited data

on assets, which limits our ability to discipline these choices in the model. The approach taken here

does could capture some aspects of risk aversion, among other factors, through heterogeneous non-

pecuniary benefits from entrepreneurship and from the utility costs of entering self-employment.

Modeling risk aversion directly awaits future work.

The expected flow utility from choosing the paid sector is

u(dit = 0, Sit, ε

0it

)= β1E [Wit|Sit] + ε0it, (1)

12

where Sit summarizes the individual’s employment history, sectoral experience, age, and beliefs

about entrepreneurial ability at time t, Wit is consumption from wage earnings in the paid sector

and ε0it is a transitory taste shock for choosing paid work that is unobserved to the econometrician.

The parameter β1 translates earnings into units of utility, scaled relative to the variance of the taste

shock. As a normalization, the expected utility of working in the paid sector for a wage of zero

is set to E(ε0it)

. To keep the model simple, we abstract from savings, eliminating intertemporal

reasons that wage earnings do not equal consumption.19 Abstracting from savings is innocuous

under risk neutrality.

The expected flow utility from choosing entrepreneurship is

u(dit = 1, Sit, β0i, ε

1it

)= β0i+β1E [Rit|Sit]+β2 (dit−1 = 0)h (Abilityi)+β3 (EntrepEx = 0)h (Abilityi)+ε

1it.

(2)

The components of utility in entrepreneurship include an unobserved taste or opportunity shock,

ε1it, along with utility over earnings plus startup costs and non-pecuniary benefits. Specifically, Rit

is consumption which equals earnings in entrepreneurship, again scaled by β1. The parameter β0i

captures non-pecuniary benefits from entrepreneurship that are specific to individual i. Individuals

with β0i > 0 are, all else equal, willing to give up some earnings to remain in self-employment; for

β0i < 0, individuals would need to earn more in self-employment than their projected paid earnings

to remain self-employed. The parameter β2 captures entry costs from moving into entrepreneurship.

Similar to Evans and Jovanovich (1989) who model constraints on borrowing as a function of wealth,

we assume that entry costs are proportional to a function of ability (meant to capture a budget

constraint based on lifetime earnings) and insert these proportional costs as a reduced form in the

flow payoffs. More will be said about this function when the state variables are introduced. While

19We also abstract from taxes. Measuring effective taxes for business owners is difficult in the PSID; however, thisissue is unlikely to have a major effect on the estimates. The parameter β1 is identified based on within-individualdifferences in expected earnings across sectors, up to scale. To the extent that differences in earnings across sectorsdo not cause drastic changes in the tax rate, interpreting the scale of β1 becomes difficult but there are no spillovereffects for the other parameters. Finally, note that the effective tax rates for most small business owners is verysimilar to individual income taxes; many small business are taxed as pass-through entities, equalizing tax schedulesacross paid work and self-employment. Businesses set up as C corporations will face a different tax schedule on anyearnings that are not distributed as salary; these earnings are subject to the corporate tax rate, and distributionsare subject to sharehold taxes.

13

these are modeled as constant proportions of functions of ability, because of risk neutrality, one can

conceptualize them as the expected costs of many factors, which may include capital investments,

effort costs of creating a new venture, the lost earnings associated with changing jobs, or the ex-

ante psychological costs of uncertainty around a new business.20 For first-time entrepreneurs, β3

captures an initial startup cost. The startup costs may differ from the entry costs to allow for

potential learning by doing in opening a business or other factors that change with experience in

selecting entrepreneurship.

Person i chooses to work in the sector that maximizes the present value of expected utility,

Vit (Sit, β0i, εit) = maxdi

E

[Ti∑τ=t

δτ−tu (diτ , Siτ , β0i, εiτ ) |Sit

], (3)

where δ is the discount rate and di is the state-contingent sequence of choices made by the individual.

We specify a flexible parametric model for earnings in each sector. If the agent is employed in

the paid sector, his earnings depend on his accumulated work experience in the paid sector, xWit,

and in entrepreneurship, xRit, along with fixed individual earnings ability in the paid sector, αi, a

transitory shock Mt, and a log AR(1) persistent shock, Pt = Pφt−1ζt. Paid earnings are given by

Wit = exp (αi +GW (xWit, xRit))PitMit. (4)

The shocks ζit andMit are distributed log-normally, with ln ζit ∼ N(

0, σ2ζ

)and lnMit ∼ N

(0, σ2

M

).

Individual earnings ability is also drawn from a log normal distribution, αi ∼ N(µα, σ

2α

), and we

assume that individuals know this ability with certainty at the time they enter the model. While

this assumption implies an asymmetry in information between sectors, it can be reframed as an

assumption that workers have had sufficient informal work experience prior to entering the paid

sector to deduce their ability with very little uncertainty.

20Evans and Jovanovic (1989) write the startup cost constraint as a function of lifetime wealth. The startup costhere is a function of expected earnings in the paid sector, a proxy for lifetime wealth.

14

Expected earnings in the paid sector in period t are therefore

E [Wit|Sit] = exp

[αi +GW (xWit, xRit) + φ log (Pt−1) +

σ2ζ + σ2

M

2

](5)

whereσ2ζ+σ

2M

2 is the convexity adjustment from the first moment of the log normal distribution.

The state variables necessary to calculate this expectation are lagged years of experience in the

paid sector and entrepreneurship, αi, and the lagged value of the persistent shock.

We assume that Pit continues to depreciate during periods when agents work in self-employment,

but does not experience any new innovations. The persistent shock in the paid sector therefore

influences moves in and out of entrepreneurship as agents with low Pit will find self-employment

temporarily more attractive.

Earnings for entrepreneurs are described by

Rit = exp [ηi +GR (xWit, xRit)] ξit, (6)

where ξit is a log-normally distributed transitory shock, ln (ξit) ∼ N(

0, σ2ξ

)and ηi ∼ N

(µη, σ

2η

)is

entrepreneurial ability. Expected earnings in self-employment differ from earnings in paid work in

several respects. Most importantly, while agents are assumed to know αi with certainty, they know

only the distribution of ηi given their paid ability but not their own endowment of entrepreneurial

ability. Initial belief about ηi are based on αi, but entrepreneurial experience is necessary to refine

beliefs about ηi. Those who enter entrepreneurship then update this belief based on their observed

earnings in that sector. In addition, earnings in entrepreneurship depend on only a transitory

shock, with no persistent stochastic element. This assumption is partially practical; a persistent

shock would substantially complicate the process by which agents update their beliefs about ηi. We

also interpret the persistent shock in the paid sector as capturing employment shocks, which are

less relevant in self-employment. Finally, accumulated experience affects earnings in each sector

differently, through the function GR (·) rather than GW (·).

To reiterate, expected earnings in entrepreneurship depend on the beliefs about entrepreneurial

15

ability. For individuals with no entrepreneurial experience, their prior belief is based on αi. To

pin down this relationship, αi and ηi are assumed to have a bivariate normal distribution with

correlation ρ; this form fits the data well and provides a tractable way to calculate the conditional

distribution of ηi given αi. This conditional prior is thus normally distributed with mean

ηi0 = µη +σησαρ (αi − µα) (7)

and variance σ2η0 = σ2

η

(1− ρ2

). For individuals with xRit years of entrepreneurial experience, the

mean belief is denoted ηix and is updated according to Bayes’ rule. This yields

ηix =σ2ξ ηi0 + xRitσ

2η0log

(Rit−1

)xRitσ2

η0 + σ2ξ

(8)

where log(Rit−1

)is the mean of the residual log earnings history in entrepreneurship from experi-

ence levels 0 through xRit−1, net of the experience profile in entrepreneurship, GR (·). The variance

of the prior distribution is updated in a deterministic fashion in each period. For xRit > 0, the

variance of the prior is σ2ηix

=σ2η0×σ

2ξ

xRitσ2η0+σ

2ξ, which declines with experience in entrepreneurship.

The risk-neutral agents in this problem care about the level of earnings. Expected earnings in

entrepreneurship are given by

E (Rit) = exp

[ηix +GR (xWit, xRit) +

σ2ηix

+ σ2ξ

2

]. (9)

As is clear from equation 9, the expected value of entrepreneurial earnings is increasing in σ2ηix

.

Holding fixed η, the expectation changes with entrepreneurial experience through σ2ηx (decreases

the convexity adjustment) and through a deterministic experience profile GR (xWit, xRit) . That is,

beliefs are a martingale and GR (xWit, xRit) is a deterministic function, so the reduction in variance

that comes with entrepreneurial experience reduces the expected flow-pay from entrepreneurship

for a fixed experience level.

16

5 Model Solution and Estimation

The previous section describes a dynamic discrete choice model with unobserved heterogeneity in

the utility parameters. In this model, non-parametric point identification of β0i (up to the other

parametric assumptions imposed on the model) is not possible for the majority of workers who never

take up entrepreneurship. We can bound the preferences of these workers from above by identifying

the highest β0i consistent with choosing the paid sector in all observed periods, but any preference

below this threshold is equally consistent with the data. This feature of our problem rules out some

highly flexible approaches to estimating the distribution of β0i, such as a finite mixture model.

Instead, we impose the distributional assumption that β0i ∼ N(µβ0

, σ2β0

). This added discipline

rules out thick tails in the preference distribution. In essence, we assume the existence of a marginal

entrepreneur.

To estimate this model we must first define the expected value to workers of choosing to work

in each sector. Because choices today affect future payoffs, the alternative-specific, or conditional,

value functions include a flow-utility term and a continuation value. These conditional value func-

tions describe the present value of the agents’ problem at time t conditional on choosing sector

dt and then following an optimal strategy in the future. The lifetime maximization problem in

equation (3) can be rewritten as a sequence of single-period decisions using the Bellman equation,

V (St, β0i, εt) = maxdt∈{0,1}

{u (dt, St, β0i, εt) + δE [V (St+1, β0i, εt+1) | dt, St, β0i]} . (10)

The value of arriving at time t with state variables St, preferences β0i, and shocks εt is the maximum

of the conditional value functions. Recall that εt is a vector of sector-specific taste shocks that are

iid across sectors and over time from distribution g(εt). Define f (St+1|St, dt; θ) as the transition

density function describing the evolution of the observed state variables, parameterized by θ. If β0i

17

is known, the conditional value function is therefore

v (dt, St, β0i, εt;β) = u (dt, St, β0i, εt;β) + δE [V (St+1, β0i, εt+1) |dt, St, β0i] (11)

= u (dt, St, β0i, εt;β) + δ

∫∫V (St+1, β0i, εt+1) g (εt+1) dεt+1f (St+1|St, dt; θ) dSt+1.

The flow utilities from choosing each sector, described in equations (1) and (2), depend on the

common parameters β and the heterogeneously distributed parameter β0i.

Given β0i, the probability of observing dt = j, where j ∈ {0, 1} indexes labor sector choice, is

given by integrating over the unobserved taste shocks εt

pt (dt = j | St, β0i;β) =

∫I

{arg maxdt∈{0,1}

v (dt, St, β0i, εt;β) = dj

}g (εt) dεt. (12)

The individual’s likelihood contribution at time t is

Lt (dt, St+1|St, β0i; θ, β) = pt (dt|St, β0i;β) f (St+1|dt, St; θ) . (13)

Since the individual’s fixed preferences are also unobserved we must instead consider the marginal

likelihood of workers’ choices given the distribution of β0i. Using this approach to recover parame-

ters, we maximize

(θ, β, µβ0

, σ2β0

)= arg max

θ,β,π

N∑i=1

log

[∫ Ti∏t=1

Lt (dit, Sit+1 | Sit, βi0;β, θ) dφ

(β0i − µβ0

σβ0

)]. (14)

The parameters to be estimated are θ, which describe the determinants of earnings and the transi-

tions of the observed state variables, β, the parameters of the flow utility functions, and µβ0, σ2β0

,

the parameters of the distribution of non-pecuniary benefits in entrepreneurship. From equation

(11), the likelihood for individual i at time t depends on the continuation value of each choice, so

each calculation of the likelihood involves solving the full lifecycle model. Estimating some param-

eters in a first step eases the computational burden by minimizing the number of parameters that

must be estimated within the full solution.

18

5.1 Determinants of Earnings and State Variable Transitions

The set of state variables, Sit, includes age, accumulated experience in each sector, xitR, xitW ,

paid sector ability, αi, expected entrepreneurial skill, ηix, the lag of the persistent shock in the

paid sector, Pit, and the lagged sector choice, dit−1. Within the model, age and experience evolve

deterministically based on sectoral choices, αi is fixed, and ηix and Pit, depend only on sector choice

and exogenous shocks. Because none of these processes depend on individually heterogeneous

parameters, when we substitute equation (13) into (14), f (St+1|dt, St; θ) moves outside of the

integral over βi and is additively separable in the log likelihood. We can therefore estimate θ

separately in a first stage, with appropriate controls for non-random selection into each sector.21

To estimate the effects of sector experience on earnings we must account for workers’ strategic

transitions between sectors. Depending on the joint distribution of abilities in each sector, tran-

sitions from paid work to self-employment may be positively or negatively correlated with paid

ability, α. We estimate earnings equation 4 including individual fixed effects, which produces con-

sistent estimates of the effect of sector experience on paid earnings, GW (·), even with non-random

selection out of the paid sector.

The effect of entrepreneurial experience on paid sector earnings is difficult to estimate for several

reasons. Both selection into entrepreneurship and moves back out of entrepreneurship are likely to

be correlated with paid sector ability and shocks to paid sector earnings. Individual fixed effects

control for some but not all of these potential biases. For example, if workers sometimes know

the realization of shocks before choosing their sector, say because they are offered a particularly

attractive new job, then workers who return to the paid sector will have disproportionately high paid

earnings. We experimented with multiple specifications for estimating this relationship. All yielded

imprecise and generally quite small estimated effects. In the current specification, we impose that

entrepreneurial experience has no effect on paid sector earnings. This zero relationship suggests that

experimenting with entrepreneurship involves an opportunity cost of lost paid sector experience,

but no additional costs or benefits. This specification of the paid sector earnings process does a

21This approach rules out some potentially interesting extensions. For example, if xitR and xitW represent ac-cumulated skills rather than years of experience, individuals could have heterogeneous learning abilities. However,with the relatively small sample in the PSID, it would be difficult to allow additional flexibility.

19

good job of fitting earnings for workers who are newly returned to the paid sector from spells of

entrepreneurship.

After netting out the experience profile, residual paid earnings consist of fixed paid ability,

the persistent wage shock log(Pt) = φlog(Pt−1) + ζt, and the transitory shock log(Mt). We set the

variance of the transitory shock in the paid sector equal to Pischke (1995)’s estimate of measurement

error in the PSID. With this calibration, the remaining parameters are identified from the moments

var(∆wt+1) = (1− φ)2σ2α + σ2

ζ +(1 + φ2

)σ2m

cov(∆wt+1, wt) = (1− φ)σ2α − φσ2

m

cov(∆wt+1,∆wt) = (1− φ)2σ2α − φσ2

m,

(15)

where ∆wt = wt − φwt−1.

These moments provide a consistent estimate of the variance of paid ability, σ2α. In the sec-

ond stage, the conditional value functions depend on each individual’s paid ability, αi. Average

residual earnings in the paid sector are an unbiased estimator of individual paid ability, but they

are inconsistent in short panels, assuming fixed-T asymptotics. With only a few years of paid

earnings, we cannot distinguish an individual with low paid ability from an individual with a low

persistent shock. Two approaches are taken to deal with these difficulties. We adjust for time-series

dependence by re-weighting residuals within individuals to place more weight on years that are less

correlated with other observations. Second, following Lazear et al. (2015), we also shrink predic-

tions αi towards the consistently estimated cell-specific means using an empirical Bayes approach.

The resulting estimator is

αi = wz + σ2α

(Σ + σ2

α

)−1(wit − wz) , (16)

where Σ is the variance-covariance matrix of the residuals using the consistent estimates of σ2α and

σ2ζ ; wz is the within-cell average residual from the paid wage equation.22 We use the average αi

across individuals in the paid sector as the measure of average paid sector ability, µα. This estimate

of µα is unbiased if there is no selection into being observed at least once in the paid sector and is

22We take means within bins by race/ethnicity and years of completed schooling.

20

consistent with large N asymptotics.

Experience-earnings profiles in entrepreneurship are subject to the same selection concerns as

profiles in the paid sector. In entrepreneurship, the bias is a direct consequence of the learning

model: workers refine their beliefs about entrepreneurial ability through observing earnings in en-

trepreneurship. As beliefs become more precise, workers who are now confident they could earn more

in the paid sector are more likely to return to that sector. This selection out of entrepreneurship

will overstate the effect of entrepreneurial experience on entrepreneurial earnings. Because most

workers have only one spell in entrepreneurship, during which paid experience remains constant,

we cannot estimate the effect of paid experience on entrepreneurial earnings within a fixed effect

model. Instead, we account for selection out of entrepreneurship by instrumenting entrepreneurial

experience with individually de-meaned experience, as in Altonji and Williams (2005).

A second source of potential bias involves selection into entrepreneurship. Workers who wait to

enter entrepreneurship, and therefore enter with more paid experience, are more likely to have been

pushed into self-employment by negative shocks in the paid sector. If this selection is correlated

with entrepreneurial performance, then estimates of the effect of paid experience on entrepreneurial

earnings will also be biased (probably downwards). To control for this selection, we include the

worker’s persistent paid-sector earnings shock in the last period before entering entrepreneurship

in the entrepreneurial earnings equation.23 We also include a set of indicator variables for race and

completed schooling. These covariates improve the precision of our experience profile estimates,

but they all capture elements of the entrepreneurial fixed effect, ηi. To estimate the stochastic

components of entrepreneurial earnings we construct residual earnings, rt by netting out only the

effects of the experience profile, leaving any heterogeneity captured by these other covariates as part

of the residual. The following set of moment conditions identify the variance components governing

23This shock is estimated in the paid earnings equation and does not depend on the experience-earnings profile inentrepreneurship. Omitting this control variable leads us to estimate substantial negative effects of paid experienceon entrepreneurial earnings.

21

entrepreneurial earnings

cov(rt+1, rt) = σ2η

cov(rt+1 − rt, rt − rt−1) = −σ2ξ .

(17)

In the final step of this stage, we estimate mean entrepreneurial ability, µη, and the correlation

between abilities in each sector, ρ. These estimates describe the transferability of skills across

sectors and allow us to recover ηi0 and σ2ηix. For workers who we observed in entrepreneurship, we

construct ηi in the same way as αi. The set of entrepreneurs may not reflect the full distribution

of entrepreneurial abilities. To account for this selection, we estimate µη as µα plus the average

difference between αi and η for individuals who we observe in both sectors. We use this same sample

of workers who appear in both sectors to estimate ρ, the correlation between α and η. We construct

ηix, the worker’s belief about his η with x years of entrepreneurial experience, using equations (7)

and (8). Finally, we construct αi for workers observed only in the entrepreneurial sector from ηi by

inverting equation (7).

5.2 Estimates of Flow Payoff Parameters

In the second stage, we take θ as given and maximize the likelihood function to estimate β, µβ0,

and σ2β0

. We follow Rust (1987) and assume that the shocks ε0it and ε1it are serially independent

with Type-1 extreme value distributions. This gives a conditional logit form for the conditional

value function

v (dt, St, β0i, εt;β) = u (dt, St, β0i, εt) + δEt [max {v (0, St+1, β0i, εt+1;β) , v (1, St+1, β0i, εt+1;β)}] (18)

= u (dt, St, β0i, εt) + δ

∫log

∑j

exp [v (dt+1, St+1, β0i, εt+1;β)]

df (St+1|St, dt; θ) + δγ,

where γ is Euler’s constant. We use these conditional value functions, equation (12), and estimates

of θ to maximize the likelihood in equation (14) for β, µβ0, and σ2

β0. We set the discount rate at

δ = 11.10 . The integral is over next year’s stochastic state variables: the persistent shock in the paid

22

sector and the belief of entrepreneurial ability. We approximate both distributions using 5-point

Gaussian quadrature.

The likelihood of the choices is then given by the conditional logit formula, pt (dt = 1 | St, β0i;β) =

exp(v(1,St,β0i;β))exp(v(1,St,β0i;β))+exp((v(0,St,β0i;β)))

. To get the marginal likelihood for sequences of choices, we inte-

grate over the distribution of β0i,

L(di|β, θ, µβ0

, σ2β0

)=

∫ T∏t=1

Lt (dit | Sit, βi0;β, θ) dφ

(β0i − µβ0

σβ0

)(19)

=

∫ T∏t=1

pt (dt = 1 | St, β0i;β)dt=1

(1− pt (dt = 1 | St, β0i;β))dt=0

dφ

(β0i − µβ0

σβ0

).

Again, we approximate the integral over β0i using 5-point Gauss-Hermite quadrature. We then

maximize∑Ni=1 ωi log (L (di|β, θ)) , the weighted log likelihood, with respect to β, µβ0

, and σ2β0.

The weights, ωi, are the average sampling weights for each individual in the PSID. Finally, because

the panel is unbalanced, individuals with longer histories are given more ”leverage.” This influences

weighting and has the potential to create deviations from the PSID sampling weights. Because of

this, we balance-adjust the panel by filling in Lmax(Ti)−Ti for those individuals who have shorter

histories than the longest-lived sample member. This does not affect individuals’ payoffs or observed

data but provides equal leverage across individuals in the sample.24

6 Estimates

6.1 Determinants of Earnings

The estimates from the first stage are presented in Table 5. Earnings in the paid sector rise

by 7% on average with the first year of paid experience and 50% with ten years of experience.

Accumulated paid experience is also associated with higher earnings in self-employment, but the

24The final likelihood to be maximized is(θ, β, µβ0 , σ

2β0

)= arg max

θ,β,π

N∑i=1

log

∫ Ti∏t=1

Lt (dit, Sit+1 | Sit, βi0;β, θ)(L)max(Ti)−Ti dφ

(β0i − µβ0

σβ0

) .

23

effects are smaller. A worker who accumulated ten years of paid experience before entering self-

employment earns 28% more on average in self-employment than a worker who entered with no paid

experience. Entrepreneurial earnings also rise with entrepreneurial experience, but not as rapidly as

paid earnings rise with paid experience. Entrepreneurs with ten years of self-employment experience

earn 25% more than new entrepreneurs on average.

The next panel of Table 5 describes the distributions of earnings abilities in each sector. The

mean ability in entrepreneurship is slightly lower than the mean paid ability. The average new

paid worker earns about $30,600 per year while the average entrepreneur with no experience earns

$29,300. We present these distributions for log weekly earnings fixed effects, so the gap corresponds

to a 4.3% drop in earnings when moving to entrepreneurship for the mean worker with no experience

in either sector. Workers who have worked in the paid sector will face larger average drops in

earnings upon entering entrepreneurship because of the smaller returns to paid experience.

Ability in entrepreneurship has more than twice the variance of ability in the paid sector,

0.61 compared to 0.21. Abilities in the two sectors are strongly correlated, so the variance of a

worker’s belief about his own entrepreneurial ability is smaller than the population variance of

entrepreneurial ability. With no experience in entrepreneurship, the variance of an individual’s

belief of his entrepreneurial ability is σ2η

(1− ρ2

)= 0.33. The relative variance of entrepreneurial

ability and the transitory shock to entrepreneurial earnings, the last panel of Table 5, imply that

this uncertainty decreases fairly quickly as workers gain entrepreneurial experience. The variance of

a worker’s belief about his ability falls to 0.07 after observing one year of entrepreneurial earnings

and 0.02 after five years. This tight correlation indicates that workers have a good prediction of

their place in the distribution of entrepreneurs, but not about their earnings potential relative to

the option of remaining in the paid sector, which is the relevant comparison for choosing where to

work.

The bottom two panels of Table 5 present the stochastic parameter estimates from the first

stage. We estimate that the persistent shock in the paid sector is far from a random walk; only 40%

of the shock persists after a year. The transitory shock to earnings in entrepreneurship is larger

than the combined shocks in the paid sector, but the real difference in uncertainty comes from the

24

variance of persistent ability.

6.2 Determinants of Utility

Table 6 presents the flow payoff parameters estimated in the second stage. The first column presents

our estimates from our baseline model with heterogeneous tastes for entrepreneurship. The scale of

utility is normalized by the taste shocks, which have a standard deviation of π√6. β1, the coefficient

on earnings in the flow utility function, converts this scale to 2010 dollars. The estimate implies

that the standard deviation of the taste shocks is large, equivalent to receiving about an additional

$105,000 in our baseline model.

The next two parameters describe the cost of entering entrepreneurship, which is allowed to differ

for former entrepreneurs and workers entering entrepreneurship for the first time. β3 describes the

start-up costs for workers making their second or more trip into entrepreneurship. Scaled by β1,

the utility costs of entering entrepreneurship are equivalent to paying about $170,000. First-time

entrepreneurs face a total cost of entering entrepreneurship equivalent to $200,000. In our model,

these utility entry costs may capture a range of financial and psychic factors. First, they may

capture real initial investments in businesses. As shown in Table 3, these initial investments vary

considerably across self-employed workers and are highly skewed. The median entrepreneur invests

only $5,400 over the first five years in self-employment, but 34% invest nothing and 10% invest

more than $65,000. In addition to direct investment, entering entrepreneurship may have real costs

in terms of foregone earnings during the transition. Finally, these utility costs may capture the

stress and pressure of developing a new business or, since we do not explicitly model risk-aversion,

a distaste for the uncertainty about entrepreneurial ability.

The final two parameters in the first column describe the distribution of non-pecuniary bene-

fits from working in entrepreneurship. The mean of this distribution, scaled by β1, implies that

individuals, on average, would accept $9,700 less in annual paid earnings to avoid working as an

entrepreneur. The standard deviation of this preference is large, about $16,000. These estimates

suggest that while the average worker dislikes entrepreneurship, 27% of workers have positive non-

pecuniary benefits and would accept lower earnings to work in entrepreneurship.

25

These parameters describe the population distribution of preferences for entrepreneurship. For

simulations and assessment of model fit, we would also like to know where in this distribution

each individual in the PSID sample is likely to fall. Define h(β0i|di, Si;µβ0

, σ2β0

)as the posterior

probability density of preferences for individual i, given their history of choices, di, and sequence

of state variables, Si. As derived in Train (2009) using Bayes’ rule, this posterior distribution is

determined by

h(β0i|di, Si;µβ0

, σ2β0

)=L (di|β, θ, β0i)φ

(β0i−µβ0σβ0

)L(di|β, θ, µβ0

, σ2β0

) , (20)

where L (di|β, θ, β0i) =∏Tt=1 Lt (dit | Sit, βi0;β, θ) is the probability of a sequence of choices con-

ditional on having preference β0i, φ(β0i−µβ0σβ0

)is the population probability distribution of that

preference, and L(di|β, θ, µβ0

, σ2β0

)is the unconditional probability of those choices as defined in

Equation (19).

The mean and variance of the individual posterior distributions are given by

µβ0i =

∫β0ih

(β0i|di, Si;µβ0 , σ

2β0

)dβ0i =

∫β0iL (di|β, θ, β0i)φ

(β0i−µβ0σβ0

)dβ0i

L(di|β, θ, µβ0 , σ

2β0

) (21)

and

σ2β0i

=

∫(β0i − µβ0i

)2 L (di|β, θ, β0i)φ

(β0i−µβ0σβ0

)dβ0i

L(di|β, θ, µβ0 , σ

2β0

) . (22)

We can then calculate the individual posterior likelihoods using Equation (19) and substituting these

individual posterior distributions in for the population distribution of preferences. The average of

these posterior mean preferences is quite close to the population average, as one would expect.

The average standard deviation of these posterior distributions is 0.035, compared to 0.039 for the

population distribution. We expect average variance of the posterior distributions to be smaller

than the population variance. The small gap indicates that considerable uncertainty remains about

individual preferences even after conditioning on the history of observed choices.

The second column of Table 6 presents the parameter estimates from an alternative model

shutting down this preference heterogeneity. Under this homogeneity assumption, we estimate

26

a smaller distaste for entrepreneurship; the fixed disutility from working as an entrepreneur is

equivalent to earning $2,100 less per year. The value of money, β1 is lower, implying an even large

role for taste shocks, and the estimated entry costs are also higher: $212,000 for repeat entrepreneurs

and $246,000 for new entrepreneurs. These changes are driven by two patterns in the data. While

exits from entrepreneurship are negatively correlated with relative entrepreneurial earnings, most

individuals who persist in self-employment earn only modestly more as entrepreneurs than they

would in paid work. This choice is incompatible with having a strong distaste for entrepreneurship.

When all workers are constrained to have the same preferences, this pattern pushes the disutility

from working in entrepreneurship towards zero. Meanwhile, for paid workers, the high variance of

entrepreneurial ability generates a strong incentive to experiment with entrepreneurship; although

the probability of being far in the right tail of entrepreneurial ability is low, the payoff to discovering

that talent is high. Nonetheless, the majority of paid workers never enter entrepreneurship. Without

heterogeneous preferences, only very high entry costs can justify this behavior. As we will see in

the next section, this simpler model cannot predict movements between sectors as well as the full

model with heterogeneity.

6.3 Model Fit

Unobserved heterogeneity and taste shocks play important roles in determining sector choices at

the individual level, particularly moves into entrepreneurship. We now assess the ability of the

full model to predict which workers will select into each sector and compare that performance

with two simpler models. Table 7 presents the average model-predicted probability of choosing

entrepreneurship, separately for workers who worked in each sector last year and who we observe

choosing each sector this year. Recall that in each interview the PSID asks about earnings in the

last calendar year. This table uses sector choice in the last calendar year and the state variables

estimated from earnings in the last calendar year to predict sector choice this year. We then check

these predictions against workers’ actual sector choice at the time of the interview.

Within each cell, the first row shows the prediction of a simpler dynamic model where workers

are forward-looking income maximizers without any preference for either sector or entry costs. Ex-

27

pected lifetime earnings are projected imposing sector choice this year and then assuming workers

make the lifetime income-maximizing choice in all future periods. With no taste shocks, each indi-

vidual has a predicted probability of choosing entrepreneurship of either 1, if the expected lifetime

earnings stream conditional on choosing entrepreneurship this period exceeds the expected stream

from choosing paid work, or 0. Income maximization predicts some of the observed variation in sec-

tor choice; among workers who were in the paid sector last year, 45% of workers who actually choose

entrepreneurship this year have higher projected lifetime earnings from that choice, compared to

only 27% of workers who choose the paid sector this year. For workers who were entrepreneurs last

year, 71% of workers who continue in entrepreneurship have higher projected earnings from that

choice compared to 45% of workers who return to the paid sector. While income maximization cor-

rectly predicts that those workers who choose entrepreneurship are more likely to do so, this simpler

model predicts far too much entrepreneurship overall. On average, an income-maximizing model

predicts that 28% of workers currently in the paid sector should choose to become entrepreneurs

next period, when in fact only 1.8% do.

The next row in each cell presents the probability of entrepreneurship predicted by our utility-

maximizing model with non-pecuniary benefits and entry costs, but without any heterogeneity in

preferences. We construct these predictions using the parameters reported in the second column of

Table 6. Under this model, high entry costs push down the share of paid workers who are predicted

to move into entrepreneurship, but they do so for everyone. Workers who choose to move from the

paid sector to entrepreneurship this period have only a 3.3% predicted probability of doing so, not

much higher than the 1.9% average predicted probability for those workers who remain in the paid

sector.

The third row reports the likelihood of selecting entrepreneurship using the estimates from our

full model with heterogeneous tastes, but using the population distribution of preferences for all

individuals. This third row represents our best estimates of individuals’ propensity to select en-

trepreneurship each year using only their characteristics (summarized by their paid earnings ability)

and work history up to that point. These predictions capture some of the variation in choices; we

predict that workers who move from paid work into entrepreneurship were about 75% more likely

28

to choose entrepreneurship while workers who leave entrepreneurship to return to paid were were

60% more likely to select paid than those workers who remained in entrepreneurship. As with the

predictions from the model with uniform tastes, these estimates improve on the predictions from a

pure income-maximizing model in terms of the overall share of workers selecting entrepreneurship.

However, these estimates use no more information on individual characteristics than the income-

maximizing model, and therefore do about equally well at separating the sector movers from the

stayers.

The final row of each cell presents the predicted likelihood of selecting entrepreneurship using the

estimates of our full model and the individual posterior distributions of tastes for entrepreneurship

described in the previous section. These estimates exploit differences in observed characteristics

across workers and also differences in unobserved tastes, as revealed by workers’ history of choices.

Keep in mind that we use each individual’s full sequence of sector choices to estimate their posterior

preferences, so the predicted likelihood of choosing entrepreneurship in year t is partially determined

by the true future choices of each worker. This row should therefore not be interpreted as a test of ex

ante predictive power, but rather as a measure of how well the two key features of our fairly simple

model, time-invariant individual tastes for entrepreneurship and time-varying individual beliefs of

entrepreneurial ability, can match observed choices. Incorporating these predicted individual tastes

improves the model’s predictive power considerably. We estimate that workers who move into

entrepreneurship are 2.6 times more likely to make that choice than workers who remain in the

paid sector. Workers who exit back into paid work are also more than twice as likely to make that

choice as workers who remain in entrepreneurship. On the entry margin, this finding is consistent

with Levine and Rubinstein (2013), who find that entrepreneurs look similar to paid workers in

terms of education, aptitude, and family background. They find that noncognitive measures have

additional predictive power to identify future entrepreneurs. Those traits will be part of what we

capture in our individual tastes for entrepreneurship.

The differences between these predictions and the earlier rows indicate that fixed unobserved

tastes play an important role in determining workers’ sectoral choices. However, the exact timing

of transitions remains difficult to predict even after accounting for fixed unobserved tastes. Our full

29

model assigns workers who we observe moving into entrepreneurship a 4.5% probability of making

that choice. This prediction is larger than the 1.7% average probability for workers who do not move

into self employment, but far below 100%. These remaining prediction errors suggest an important

residual role for time-varying shocks. We characterize these iid shocks as taste shocks in our model,

but they may also capture other time-varying unobserved states. For example, they may represent

the arrival of an innovative idea or access to capital on the entry margin, or bankruptcy or other

business destruction on the exit margin.

The model does a particularly good job of matching the rate of transitions in and out of en-

trepreneurship over the lifecycle. Figure 4 plots the probability of moving from the paid sector

into entrepreneurship by age. The data series plots the share of individuals who worked in the

paid sector in the previous year who choose to enter self-employment in the current year, while the

model series plots the average predicted likelihood of selecting entrepreneurship for these workers,

using the posterior taste distributions. The predicted likelihoods match the broad pattern of the

data: entry rates are higher for younger workers, averaging around 2% for workers in their 20s, and

fall as workers age, to an average of less than 1% for workers over 45. The fall in entry rates later

in life reflects two factors. First, paid experience increases earnings in the paid sector more than it

increases earnings in entrepreneurship, so more experienced workers are less likely to improve their

earnings by entering entrepreneurship. Second, the option value of learning one’s entrepreneurial

ability is highest for younger workers, who have more remaining working years to make use of that

information. In the data, the probability of becoming self-employed rises unsteadily through the

20s and early 30s, peaking at 3% at age 33, before falling for the rest of the lifecycle. The mean

model-predicted likelihood of choosing self-employment peaks at 2.6% for the youngest workers and

then falls steadily with age. The observed rise in moves to self-employment through the 20s may

reflect credit constraints, which we do not model explicitly.

Figure 5 plots the probability of returning to the paid sector by years in self-employment. In our

model, as in the data, the hazard rate of leaving entrepreneurship falls sharply with entrepreneurial

experience. This decline reflects the sharp fall in uncertainty about entrepreneurial ability. With

no experience, the variance of an individual’s belief of his entrepreneurial ability is 0.33. This

30

uncertainty falls to 0.07 after a year, 0.02 after five years of entrepreneurial experience, and 0.01 after

ten years. The model predicts that nearly 20% of workers will leave entrepreneurship after 1 year,

while only 13% of workers who remain after five years will exit the following year. The empirical

hazard rates closely follow these predicted exit rates. Observed exit rates are slightly higher than

the model predicts for the first few years, which may reflect misclassification measurement error if

workers with very short measured spells in entrepreneurship were actually never entrepreneurs.

7 Decomposing the Value of Entering Entrepreneurship

Transitions out of entrepreneurship are strongly correlated with expected individual earnings in

each sector. This behavior is consistent with an option value in entrepreneurship. Workers who

enter entrepreneurship gradually learn about their entrepreneurial skills. As shown in Figure A1,

those with high relative entreprenurial earnings are more likely to remain in that sector, while those

with lower expected relative earnings are more likely to return to paid work.

7.1 Selection into Entrepreneurship

Cross-sectional differences in earnings across sectors, that is comparisons of average earnings for

workers observed in each sector, are a poor predictor of individual expected gains to entering

entrepreneurship for two reasons. First, these cross-sectional differences are biased estimates of

expected static earnings because of selection. The set of workers currently in entrepreneurship will

overweight workers with high relative entrepreneurial earnings because those workers are most likely

to persist in entrepreneurship. Table 8 quantifies this bias. The first panel of this table displays

observed earnings for 32 year olds who worked in each sector. We focus on this one age group to

eliminate the effect of different average ages across sectors and to simplify the interpretation when

we move to forward-looking projections. Consistent with Table 1, and earlier research, workers in

entrepreneurship earn more at the mean and less at the median.

The next panel of this table summarizes projected current year earnings in each sector for

all 32 year olds in the sample, based on the parameter estimates in Table 5. In the first panel,

31

workers contribute to only one column each: entrepreneurs are included in the entrepreneurial

average and paid workers in the paid average. In this second panel, all 32 year olds contribute

to both averages. The difference between these first two panels quantifies the selection bias in

cross-sectional earnings. Because successful entrepreneurs are more likely to persist, the mean

annual entrepreneurial earnings for current entrepreneurs is $5,000 higher than the mean expected

earnings in entrepreneurship across all workers. 32 year old entrepreneurs earn 26% more than 32

year old paid workers at the mean, but the average 32 year old would expect to earn only $519

more, or 0.1% of his paid earnings, if he entered entrepreneurship. The selection bias is smaller and

of the opposite sign at the median, perhaps reflecting the presence of some adverse selection into

entrepreneurship. Because entrepreneurial earnings have a long right tail, overweighting successful

entrepreneurs brings up the mean more than the median. This bias is even more pronounced if

we consider older workers, because older entrepreneurs have more experience on average and are

therefore even more heavily selected.

7.2 The Option Value of Experimentation

Both of these first two panels describe a single year of earnings. Because of the option value of

experimentation, even unbiased estimates of the single-period earnings in each sector will misstate

the expected value of entering entrepreneurship. The last two panels of Table 8 describe expected

discounted lifetime earnings, conditional on choosing each sector today. To make these numbers

comparable to one-year earnings we convert the discounted sum of lifetime earnings to annual

equivalents: the constant annual earnings that would generate the same discounted lifetime earnings

as our estimates.

The third panel describes expected lifetime earnings in a static model where workers must remain

in one sector for all future periods. Because earnings in both sectors rise with experience, annualized

projected lifetime earnings are slightly higher than projected earnings this year even without an

option value. The increase between the second and third panels is larger for entrepreneurial earnings

than for paid because of the patterns of accumulated sector experience among 32 year olds and the

concavity of returns to experience. Most of these workers have accumulated close to a decade of paid

32

experience, so their current earnings in the paid sector reflect that experience and their projected

future earnings incorporate only modest gains to additional experience. Few workers have any

entrepreneurial experience, so the one-year projected earnings in entrepreneurship are lower, but

the lifetime projections anticipate sharp earnings growth over the first several years in that sector.

The last panel of Table 8 describes projected lifetime earnings imposing sector choice today

and then assuming workers choose the sector that maximizes expected utility in all future years

as described by our full model using the posterior estimates of individual tastes. The differences

between this panel and the one above it measure the option value of entering entrepreneurship. At

the mean, the expected lifetime value of entering entrepreneurship in this dynamic model is $7,000

higher per year than the expected value of selecting into entrepreneurship in the static model.

On average, 32 year olds with no entrepreneurial experience expect to earn 11% more by entering

entrepreneurship with the option to return to the paid sector than by working as an entrepreneur

in all future years.25 In the static model, the mean expected lifetime earnings gain from entering

entrepreneurship is 1.5% of paid earnings and the median expected gain is 0.6% of paid earnings.

With the addition of the option value, the mean percent gain rises to 11% of paid earnings.

Comparing the top panel of Table 8 to the bottom panel, we see that selection bias overwhelms

the omission of the option value at the mean; the simple cross-sectional comparison suggests that

workers can expect to earn 26% more at the mean while the full model implies an average gain of

only 11%. The reverse is true for the median worker. According to the cross-sectional estimates,

the median worker should expect to lose 6.7% of his earnings if he enters entrepreneurship, while

the static model predicts approximately equal future earnings in each sector. However, the dynamic

value of entering entrepreneurship is 12% higher at the median than the value of choosing the paid

sector.

Even among workers of the same age, individual expected gains from entering entrepreneurship

vary considerably from these averages. The largest determinant of individual expected gains is

past entrepreneurial experience. Workers who have never been entrepreneurs have considerable

uncertainty about their entrepreneurial ability, and therefore a high option value of experimenting

25This is the mean percent difference across individuals, not the percent difference in the means.

33

with entrepreneurship. Once workers enter entrepreneurship that uncertainty resolves; workers

with high entrepreneurial ability relative to their paid ability continue to expect higher earnings in

entrepreneurship while those with lower revealed entrepreneurial ability expect lower earnings.

Figure 6 illustrates these patterns by plotting the difference in annualized expected lifetime

earnings conditional on choosing entrepreneurship this year instead of the paid sector. These

differences are plotted separately for workers with no entrepreneurial experience, workers who were

self-employed last year, and workers who have some entrepreneurial experience but have since

returned to the paid sector. Expected gains from entering entrepreneurship are highest for workers

who have no past entrepreneurial experience. Most workers with no entrepreneurial experience

expect positive earnings gains from entering entrepreneurship, with a mean annual gain of $9,200.

As uncertainty about entrepreneurial ability resolves, workers who discover they have a comparative

advantage in entrepreneurship remain entrepreneurs while those who learn otherwise are more

likely to return to paid work. Current entrepreneurs anticipate earning $10,100 more per year on

average if they remain in entrepreneurship than if they return to the paid sector, while workers

who have already left entrepreneurship anticipate earning -$264 less per year if they return to

entrepreneurship.

8 Conclusion

Earnings for entrepreneurs are higher than earnings for paid workers at the mean, but lower at

the median. Self-employment requires ingenuity, self-discipline, and a wide range of skills. It also

involves more income volatility than paid work, which will deter risk-averse workers, and does not

offer the kind of non-wage benefits, such as subsidized health insurance, that many paid workers

receive from their employers. For all these reasons, we would expect earnings to be higher in

entrepreneurship in order to rationalize selection in. Earlier research has used these cross-sectional

comparisons of earnings as a proxy for what individual workers could expect to earn in each sector

and proposed several solutions to the puzzlingly low median earnings in entrepreneurship.

In this paper, we show that these cross-sectional comparisons are poor estimates of what an

34

individual worker could expect to gain from entering entrepreneurship. First, the higher cross-

sectional mean earnings for entrepreneurs are entirely due to positive selection. Many workers

enter self-employment, but those with high earnings relative to their options in the paid sector are

far more likely to remain self-employed. After correcting for this selection, expected annual earnings

are about equal across the two sectors, while median earnings remain lower in entrepreneurship.

However, earnings in a single year are an incomplete measure of the value of entering en-

trepreneurship. Uncertainty about entrepreneurial ability, combined with an option to move back

and forth between the paid sector and self-employment, creates a substantial value to experiment-

ing with entrepreneurship. Workers who enter self-employment will be able to learn about their

entrepreneurial skills, and then remain in entrepreneurship only if the expected returns to that

choice exceed the expected gains from returning to the paid sector. We use a dynamic two-sector

labor model to estimate the size of this option value and find that it is substantial. The option

value adds the equivalent of $7,500 per year to the expected value of entering entrepreneurship. We

also show that low relative earnings in entrepreneurship strongly predict returns to the paid sector,

consistent with this experimental behavior. This option value is so large that, in a dynamic model,

the expected returns to entering entrepreneurship exceed expected returns in the paid sector, even

at the median. To rationalize observed patterns of sector choice, the average worker must dislike

working in entrepreneurship. However, we estimate that the non-pecuniary benefits of working

in entrepreneurship are highly heterogeneous across workers; about a quarter of workers prefer

self-employment.

35

References

Altonji, Joseph G and Nicolas Williams (2005) “Do wages rise with job seniority? A reassessment,”

Industrial & Labor Relations Review, Vol. 58, pp. 370–397.

Bruce, Donald and Herbert J Schuetze (2004) “The labor market consequences of experience in

self-employment,” Labour Economics, Vol. 11, pp. 575–598.

Caliendo, M and A Uhlendorff (2008) “Self-employment dynamics, state dependence and crossmo-

bility patterns. IZA Discussion Papers 3900,” Institute for the Study of Labor (IZA), Bonn.

Caliendo, Marco, Frank Fossen, and Alexander Kritikos (2010) “The impact of risk attitudes on

entrepreneurial survival,” Journal of Economic Behavior & Organization, Vol. 76, pp. 45–63.

Evans, David S and Boyan Jovanovic (1989) “An estimated model of entrepreneurial choice under

liquidity constraints,” The Journal of Political Economy, pp. 808–827.

Evans, David S and Linda S Leighton (1989) “Some empirical aspects of entrepreneurship,” The

American Economic Review, pp. 519–535.

Hall, Robert E and Susan E Woodward (2010) “The Burden of the Nondiversifiable Risk of En-

trepreneurship,” The American Economic Review, Vol. 100, pp. 1163–1194.

Hamilton, Barton H (2000) “Does entrepreneurship pay? An empirical analysis of the returns to

self-employment,” Journal of Political economy, Vol. 108, pp. 604–631.

Hurst, Erik, Geng Li, and Benjamin Pugsley (2014) “Are household surveys like tax forms? Evi-

dence from income underreporting of the self-employed,” Review of economics and statistics, Vol.

96, pp. 19–33.

Hurst, Erik and Benjamin Wild Pugsley (2011) “What do small businesses do?”Technical report,

National Bureau of Economic Research.

Lazear, Edward P, Kathryn L Shaw, and Christopher T Stanton (2015) “The value of bosses,”

Journal of Labor Economics, Vol. 33, p. 823861.

36

Levine, Ross and Yona Rubinstein (2013) “Smart and Illicit: Who Becomes an Entrepreneur and

Does It Pay?” working paper.

Light, Audrey and Robert Munk (2015) “Business ownership vs. self-employment,” working paper.

Manso, Gustavo (2014) “Experimentation and the returns to entrepreneurship,” working paper.

Moskowitz, Tobias J and Annette Vissing-Jorgensen (2002) “The returns of entrepreneurial invest-

ment: A private equity premium puzzle?” The American Economic Review, Vol. 92, p. 745.

Pischke, Jorn-Steffen (1995) “Measurement error and earnings dynamics: Some estimates from the

PSID validation study,” Journal of Business & Economic Statistics, Vol. 13, pp. 305–314.

Rust, John (1987) “Optimal replacement of GMC bus engines: An empirical model of Harold

Zurcher,” Econometrica: Journal of the Econometric Society, pp. 999–1033.

Train, Kenneth E (2009) Discrete choice methods with simulation: Cambridge university press.

Vereshchagina, Galina and Hugo A Hopenhayn (2009) “Risk taking by entrepreneurs,” The Amer-

ican Economic Review, pp. 1808–1830.

37

Table 1: Summary Statistics for Men in the PSID

Never Entrepreneurs Sometime EntrepreneursNumber of individuals 5,354 1,549Years spent in paid work 10.23 7.04Years spent in entrepreneurship 0.00 4.78Ever own a business 0.21 0.82Less than HS diploma 0.11 0.13High school diploma 0.40 0.35Some college 0.27 0.28College graduate 0.19 0.19Graduate degree 0.03 0.04White 0.79 0.88Black 0.12 0.05Hispanic 0.07 0.05Other non-white 0.02 0.01

Source: PSID 1976-2011. Averages use weights as described in the text. Note that business ownership while in

entrepreneurship, reported in the text, is calculated using spells in entrepreneurship.

38

Table 2: Transitions between Sectors and Incorporation Status

Entrepreneurs

Paid No Business Unincorporated Incorporated

Paid sector 97.8 0.9 0.9 0.4

Entrepreneurs, no business 18.9 39.6 27.7 13.7

Entrepreneurs, unincorporated 10.8 9.4 71.5 8.4

Entrepreneurs, incorporated 5.1 5.5 8.4 80.9

Source: PSID 1976-2011. Annual transition probabilities in an out of entrepreneurship and across business

ownership and incorporation status.

39

Table 3: Initial Capital Investments for New BusinessesShare of Mean Paid Mean Entrep. Std. deviation

Distribution of investments observations Fixed Effect, α Fixed Effect /α Ent. Effect /αNo investment 34 % 6.44 1.05 0.25$1-5,000 15 % 6.25 0.98 0.08$5,001-25,000 26 % 6.41 1.01 0.09$25,001-65,000 15 % 6.63 1.02 0.08More than $65,000 8 % 6.66 1.05 0.11

Source: PSID 1984-2011. This table describes the behavior of the 272 PSID respondents who are asked about

assets and investments in their business within 5 years of entering entrepreneurship and covers total investments

over the first 2-5 years. We include investments of $0 when individuals are asked about investing in their business

and report no investment. This measure likely overstates the financial cost of entering self-employment because

it excludes individuals who become self-employed without starting a business. Statistics are separated roughly by

quartiles of initial investment, with the top few investors separated out. The last two columns describe the ratio of

each individual’s final belief of entrepreneurial ability η, that is the belief based on the most years of entrepreneurial

earnings, to each individual’s paid ability.

40

Table 4: Determinants of Weekly Earnings

OLS 25th pctile 50th pctile 75th pctileEntrepreneur 403.2 -162.1 -42.97 258.1

(40.86) (11.65) (13.14) (27.13)Age 63.81 53.08 68.36 74.15

(6.544) (2.229) (2.530) (3.856)Age-squared -0.473 -0.517 -0.623 -0.587

(0.0941) (0.0322) (0.0359) (0.0557)Black -250.8 -196.0 -213.9 -264.4

(11.45) (6.707) (8.176) (10.82)Hispanic 3.732 -68.44 -89.87 -103.8

(25.93) (13.14) (12.07) (18.44)Other 59.63 34.31 -13.68 52.63

(64.60) (39.35) (51.82) (42.65)Less than HS -225.4 -175.6 -187.7 -204.9

(9.146) (5.922) (8.516) (10.52)Some Col 107.1 51.97 78.74 104.9

(9.893) (6.958) (6.646) (9.188)Col Grad 616.8 299.4 415.8 552.0

(22.93) (8.791) (10.00) (13.69)Grad Degree 1027.2 527.6 680.2 903.5

(45.74) (18.78) (18.36) (34.57)Mean Fitted Values 1,075 720 1,000 1,333for PaidObservations 67,228 67,228 67,228 67,228

PSID 1976-2011. Columns present OLS and quantile regressions where the dependent variable is real weeklyearnings in 2010 dollars. Standard errors in parentheses. Estimates use sampling weights from the PSID asdescribed in the text.

41

Table 5: Determinants of Log Earnings

Experience profiles in paid sector1 year paid experience 0.070

(0.003)10 years paid experience 0.492

(0.213)Experience profiles in entrepreneurship

1 year paid experience 0.046(0.035)

10 years paid experience 0.279(0.430)

1 year entrepreneurial experience 0.039(0.014)

10 years entrepreneurial experience 0.250(0.182)

Distribution of abilitiesMean log ability in paid sector µα 6.416

(0.012)Variance of log ability in paid sector σ2

α 0.213(0.009)

Mean log ability in entrepreneurship µη 6.373(0.092)

Variance of log ability in entrepreneurship σ2η 0.615

(0.048)Correlation of abilities across sectors ρ 0.679

(0.054)Paid sector earnings shocks

Variance of AR(1) innovation σ2ζ 0.048

(0.002)Annual persistence of AR(1) φ 0.417

(0.047)Variance of transitory shock σ2

m 0.010

Entrepreneurship earnings shockVariance of transitory shock σ2

e 0.094(0.008)

PSID 1976-2011. Estimated on weekly earnings as described in the text. Bootstrapped standard errors from 200 drawsin parentheses. Paid earnings depend on a cubic in paid experience. Entrepreneurial earnings depend on a quadraticin paid experience and a cubic in entrepreneurial experience. The order of these polynomials were determined by aseries of specification F-tests. We discuss omitting entrepreneurial experience from the paid earnings equation in thetext. Variance of transitory shock is calibrated to the estimate of measurement error in PSID from Pischke (1995).In simulations annual earnings are projected from these estimates assuming 50 weeks of work per year.

42

Table 6: Utility Parameters

Full model No heterogeneous tastesβ1: Utils per $10,000 0.122 0.100

(0.011) (0.004)β2: Cost of first entry -0.393 -0.348

(0.083) (0.058)β3: Startup costs -2.059 -2.121

(0.070) (0.056)µβ0

: Mean non-pecuniary benefit -0.119 -0.022(0.034) (0.004)

σβ0: Std. dev. of non-pecuniary benefit 0.198

(0.022)

The utility parameter per $10,000 has no independent interpretation except to scale the variance of the extreme-value

taste shock. See text for details of estimation.

43

Table 7: Estimated Probability of Choosing Entrepreneurship

Observed Choice this YearSector Last Year: Paid Sector EntrepreneurshipPaid sector

Income maximizing 27.3% 44.6%Full model, uniform tastes 1.9% 3.3%Full model, heterogeneous tastes 1.7% 3.0%Full model, posterior 1.7% 4.5%

Workers in cell 52,844 855Share of total workers 88.3% 1.4%

EntrepreneurshipIncome maximizing 45.3% 70.8%Full model, uniform tastes 80.7% 87.6%Full model, heterogeneous tastes 75.6% 84.5%Full model, posterior 77.7% 89.7%

Workers in cell 519 5,612Share of total workers 0.9% 9.4%

Each column presents the model-predicted probability of choosing entrepreneurship, separately for workers who

were in each sector last year and are observed to choose each sector this year. The first row in each cell gives

the prediction of a model where workers are income maximizers with no entry costs or non-pecuniary benefits of

entrepreneurship. The next row presents predictions from a model where agents are utility maximizers with our full

preference specification, but without any heterogeneity of preferences. The third row present the predictions of our

full model with preference heterogeneity.

44

Table 8: The Value of Entering Entrepreneurship at age 32

Value of Paid Value of Entrep. Difference As % of Paid ValueAnnual observed earnings in chosen sector

Mean 59,038 74,528 15,490 26.2%Median 54,562 50,909 -3,653 -6.7%

Projected annual earnings for all workersMean 58,512 59,031 519 0.1%Median 54,017 52,200 -1,817 -4.7%

Projected lifetime earnings, static modelMean 67,444 70,330 2,886 1.5%Median 62,103 62,252 150 0.6%

Projected lifetime earnings, dynamic modelMean 67,974 77,835 9,861 11.2%Median 62,083 68,613 6,530 12.2%

The first panel of this table summarizes observed annual earnings for 32 year old paid workers and entrepreneurs.

In this panel the workers contributing to the paid averages are different from the workers contributing to the

entrepreneurial average. In the remaining panels, earnings are projected for all 32 year old workers conditional on

choosing each sector, so the same sample of workers contribute to the averages for each sector. For easier

comparison, the projected lifetime earnings are converted to constant-annual-income equivalents, C such that∑Ts=t

(1

1+r

)sYs = C

∑Ts=t

(1

1+r

)s. The third panel projects discounted lifetime earnings assuming that workers

choose each sector this year and remain there for all future years. The last panel projects discounted lifetime

earnings assuming workers choose each sector this year and then behave optimally according to the full model in

each subsequent year.

45

Figure 1: Composition of Entrepreneurs Over Time0

.1.2

.3.4

.5.6

.7.8

.91

Sha

re o

f sta

rting

ent

repr

eneu

rs

1 6 11 16Years since Start of Entrepreneurship

Returned to paid sector Self-employed, no businessUnincorp. business owner Incorp. business owner

Source: PSID 1976-2011. Note: Because of bi-annual interviews after 1997 the sample is not consistent over years

since entering entrepreneurship. For example, we may see some workers 2 and 4 years post-entry and others 3 and

5 years post-entry.

46

Figure 2: Distribution of Earnings in Wage Work and Entrepreneurship

0.0

002

.000

4.0

006

.000

8D

ensi

ty

0 1000 2000 3000 4000Weekly Earnings, 2010 USD

Wage workers Entrepreneurskernel = epanechnikov, bandwidth = 54.7896

Source: PSID 1976-2011. Distribution of real weekly earnings in 2010 dollars. Truncated at $4,000 per week, which

excludes the top 2% of earnings. Weighted as described in the text.

47

Figure 3: Earning Profiles by Persistence in Entrepreneurship

3000

040

000

5000

060

000

7000

0M

edia

n An

nual

Ear

ning

s, 2

010

USD

0 2 4 6 8 10Year of Entrepreneurship

Spell lasts 6 or more years Spell lasts 2-5 yearsSpell lasts less than 2 years

Source: PSID 1976-2011. Profiles are average real annual earnings for entrepreneurs in 2010 dollars. Weighted as

described in the text. The gap between each of the two lower profiles and the top profile are statistically significant

with 99% confidence.

48

Figure 4: Entries into Entrepreneurship by Age

0.0

1.0

2.0

3P

r of M

ovin

g in

to E

ntre

pren

eurs

hip

20 25 30 35 40 45 50 55Age

Model Data

Source: PSID 1976-2011 and predicted likelihood of choosing entrepreneurship from the model. Both series describe

the average probability of choosing to work in entrepreneurship for individuals who worked in the past sector in the

previous year. Data are weighted as described in the text.

49

Figure 5: Propensity to Move to Paid Sector by Time in Entrepreneurship

0.0

5.1

.15

.2.2

5P

r of E

xitin

g E

ntre

pren

eurs

hip

0 5 10 15 20Years of Entrepreneurial Experience

Model Data

Source: PSID 1976-2011 and predicted likelihood of choosing paid work from the model. Both series describe the

probability of selecting paid work for individuals who worked as entrepreneurs in the previous year. Data are weighted

as described in the text.

50

Figure 6: Expected Lifetime Earnings Gains from Entering Entrepreneurship

0.0

0002

.000

04.0

0006

.000

08D

ensi

ty

-20000 0 20000 40000 60000Expected gains from choosing entrepreneurship this year

Current entrepreneurs Former entrepreneursNot yet entrepreneurs

Figure plots the annualized expected lifetime earnings conditional on choosing entrepreneurship this year less the an-

nualized expected lifetime earnings conditional on choosing the paid sector this year, both for 32 year olds. This is the

full distribution of the projected lifetime earnings summarized in the bottom panel of Table 8. Current entrepreneurs

are those with positive entrepreneurial experience who also select entrepreneurship this year. Former entrepreneurs

have positive entrepreneurial experience but have already returned to the paid sector. Not yet entrepreneurs have

no entrepreneurial experience, although they may select into entrepreneurship this year.

51

Figure A1: Relative Earning Profiles by Persistence in Entrepreneurship

.81

1.2

1.4

1.6

Med

ian

Rel

ativ

e En

trepr

eneu

rial E

arni

ngs

0 2 4 6 8 10Year of Entrepreneurship

Spell lasts 6 or more years Spell lasts 2-5 yearsSpell lasts less than 2 years

Source: PSID 1976-2011. Profiles are the ratio of average observed annual earnings for entrepreneurs to their

projected earnings had they worked in the paid sector that year, constructed using the estimates in Table 5 .

Weighted as described in the text.

52