Availability of the Gig Economy and

Availability of the Gig Economy and

Long Run Labor Supply Effects for the Unemployed

Emilie Jackson *

November 2022

Abstract

A growing number of American workers earn income through platforms in the gig economy

which provide access to fexible work (e.g. Uber, Lyft, TaskRabbit). This major labor market innovation presents individuals with a new set of income smoothing opportunities when

they lose their job. I use US administrative tax records to measure take up of gig employment

following unemployment spells and to evaluate the effect of working in the gig economy on

workers’ overall labor supply and earnings trajectory. To do so, I utilize penetration of gig platforms across counties over time, along with variation in individual-level predicted propensities

for gig work based on pre-unemployment characteristics. In the short run, I show an increase

in gig work following an unemployment spell and that individuals are correspondingly better

able to smooth the resulting drop in income. However, individuals stay in these positions and

are less likely to return to traditional wage jobs. Thus, several years later, prime-age (25-54)

workers’ income lags signifcantly behind comparable individuals who did not have gig work

available. Among older workers (55-69), I fnd an increase in gig work corresponds to a reduction in receipt of Social Security Disability Insurance (SSDI). I shed light on mechanisms at

play and show that either these individuals have extreme values for fexibility, and are outliers

in their preferences, or individuals are perhaps procrastinating searching for a new job and not

fully optimizing.

* Jackson:

Department of Economics, Michigan State University, East Lansing, MI 48824. Email: emiliej@msu.edu. This research and access to tax data was authorized by the Treasury Offce of Tax Analysis and the

IRS under an academic partnership program. The fndings, interpretations, and conclusions expressed in this paper are

entirely those of the authors and do not necessarily refect the views or the offcial positions of the U.S. Department of

the Treasury or the Internal Revenue Service. Any taxpayer data used in this research was kept in a secured Treasury or

IRS data repository, and all results have been reviewed to ensure no confdential information is disclosed. I would like

to thank Mark Duggan, Paul Oyer, Raj Chetty, Petra Persson, and Gopi Shah Goda for their invaluable guidance and

support throughout this project. I would also like to thank Ran Abramitzky, Nick Bloom, Liran Einav, Andy Garin,

Matt Gentzkow, Matt Jackson, Dmitri Koustas, Adam Looney, Alicia Miller, Luigi Pistaferri, Shanthi Ramnath, Isaac

Sorkin, Heidi Williams, and seminar participants at Colegio de Mexico, Clemson, IE Business School, U. of Michigan, U. of Michigan - Ford, Michigan State, Monash, Notre Dame, Princeton, Stanford, UCSB, UIUC, Brookings,

Lyft, RAND, Treasury - Offce of Tax Analysis, SEA Annual Meeting, MD4SG’20, Young Economist Symposium,

13th Annual All California Labor conference, and 14th Annual EGSC for helpful comments and conversations. I am

grateful to the many people at OTA and IRS who made this work possible. I gratefully acknowledge funding support

from Peter G. Peterson Foundation for a post-doctoral fellowship at NBER, Alfred P. Sloan Foundation Pre-doctoral

Fellowship on the Economics of an Aging Workforce, Laura and John Arnold Foundation, and B.F. Haley and E.S.

Shaw Fellowship for Economics.

1 Introduction

In the last decade, there has been a substantial increase in the number of individuals earning income

through the “gig economy,” by which I mean online platforms such as Uber or TaskRabbit.1 As

seen in Figure 1, the number of individuals with any gig work in the United States (US) increased

from less than one thousand in 2007 to almost two million by 2016. This rapid emergence of the

gig economy represents a potentially major change in the labor market, and presents workers with

more fexible work opportunities.

I focus in this paper on the fact that gig labor market opportunities may be especially relevant

for the unemployed population since they provide short-term, fexible work that permits workers to

recover some of their lost income after job loss. This provides a new and valuable form of insurance

while workers search for a job to re-enter the workforce. However, it is ultimately an empirical

question as to whether short-term work in the gig economy – which may change individuals’ job

search behavior and delay (even indefnitely) their re-entrance into traditional employment – is

benefcial in the longer term.

To investigate this issue, I quantify the take up of gig work following an unemployment spell

and estimate the causal impact of participating in gig work on a workers’ short-run and long-run

earnings, employment, and education decisions. In doing so, I evaluate the trade-off between

smoothing income in the short run and less attachment to traditional work in the long run, measuring the extent to which either or both are present. I utilize the universe of individual Federal

income tax returns for the US and follow a panel of individuals who lose their job, which I measure based on one’s receipt of unemployment insurance (UI).2 I examine earnings, job type (e.g.

gig versus wage employment), post-secondary school attendance, and social insurance receipt (e.g

Social Security Disability Insurance (SSDI) and Social Security retirement) as they evolve before,

during, and after job loss.

My empirical strategy leverages variation in gig platform availability across counties over time

within the US that is driven by the geographic rollout of gig economy platforms. This exploits

the fact that within a given county some individuals will have the additional option of working in

the gig economy depending on the year in which they lose their job (i.e. those who lost their job

after the entry of gig platforms). However, a simple difference-in-differences approach would not

be able to disentangle the effects of gig availability from local labor-market changes happening

differentially in places gig platforms entered earlier. This is potentially the case since platforms’

decisions on where to enter frst were largely driven by population, starting with highly populated

areas, and larger cities may have recovered differentially from the Great Recession during the time

1 The “gig economy” is used colloquially to refer to digital platforms that match consumers and providers. I focus on

gig employment through online platforms rather than contract work more broadly.

2 In the US, unemployment compensation is taxable income and therefore identifable in the tax data.

1

period that I examine.3

To deal with this selective gig rollout, I incorporate within-area variation in individuals’ predicted propensity for gig work by splitting individuals into two groups: high and low-gig-propensity.4

In doing so, I account for local labor-market changes that affected all individuals within a countyyear. For instance, high earners (prior to job loss) should not alter their behavior and so by accounting for the differences that they experience when gig platforms are available compared to

when they are not helps me control for the possibility that outcomes are changing differentially

in places where platforms entered earlier. I estimate a probit regression of the decision to ever

participate in gig work on numerous pre-unemployment characteristics for individuals who had

gig platforms available to them at the time of job loss. Utilizing these estimates, I identify comparable individuals who plausibly would have taken up gig work had it been available but simply

did not have it as an option. I refer to these individuals as high-gig-propensity, and the remaining

individuals as low-gig-propensity, excluding the bottom half of the propensity distribution as these

individuals are less similar.

This technique yields a high-gig-propensity group that is 18 percentage points more likely to

engage in gig work than the identifed low-gig-propensity workers, among individuals with the

most gig availability relative to no gig availability. Prior to losing their main job and receiving UI,

essentially no individuals were working in the gig economy to any degree.5 Following job loss, I

document an extensive-margin increase in gig work. Among the overall sample of UI recipients,

about 1-2% of individuals take up gig work following an unemployment shock.

Combining variation in gig availability and individuals’ propensity for gig work, I estimate

a triple-difference specifcation that estimates the impact of gig availability following job loss

(Gruber, 1994). I establish three main fndings. First, high-gig-propensity individuals with the

highest degree of gig availability experience a short-term beneft in income relative to those without

gig availability. Their individual and household income drop by $3,000 less in the year of UI

receipt than those without gig platforms available. However, the income of those without gig

availability catches up the year after UI receipt.

Second, I fnd that despite a smaller drop in income in the short run, two to four years later

the income of prime-age workers who had gig platforms available when and where they lost their

3 Population has an R squared value of 0.75 in predicting the year that gig platforms enter a county.

This R square

comes from a simple regression of the year of gig entry on 2014 county population as a cubic to pick up the curvature

of the relationship.

4 Similar methodology has been employed in other papers, e.g. Banerjee et al. (2018) use within area variation in individuals likelihood of taking up microfnance and compare individuals with high and low propensity for microfnance

in villages that do and do not receive microfnance as an option.

5 This is partly by construction as all of these individuals must have started with a traditional wage job in order to lose

it. However, as I show with co-authors in Collins et al. (2019), the majority of workers in the gig economy overall are

actually individuals who hold a main wage job and gig work is a secondary source of earnings.

2

job begins to lag signifcantly behind those who did not have gig platforms available, and this is

driven by lower wage earnings. I show that individuals without gig availability begin to return

to wage jobs, while individuals with gig availability stay in gig positions and are fve percentage

points less likely to return to traditional wage jobs two to four years after UI receipt. This translates into $4,000 lower wage earnings and household income, relative to their high-gig-propensity

counterparts without any gig platforms available.

Third, I establish important heterogeneity across ages by differentiating between prime-age

workers (ages 25-54) and older workers (ages 55-69), as older individuals have been shown to

highly value fexibility (Ameriks et al., 2017) and are especially vulnerable after losing their job.6

Crucially, the implications of entering into gig work depend on the set of relevant and available

options following job loss for each age. Empirically the outside options differ dramatically for

these two age groups. For prime-age workers, gig work crowds out wage jobs that provided higher

earnings as well as important employer-sponsored benefts. In contrast, for older workers, gig

work prolongs labor force participation. An increase in gig work instead reduces receipt of Social

Security Disability Insurance (SSDI) benefts and may delay claiming of Social Security retirement

benefts, which can be fnancially advantageous.

This study helps to distinguish among mechanisms that are consistent with the observed patterns for prime-age workers. First, it may be that individuals learn that they value fexibility after

entering into gig work. Alternatively, individuals may have time-inconsistent preferences, and plan

to search for a new job yet keep procrastinating. Estimates I fnd imply that individuals are willing

to forgo 39% of their earnings in exchange for fexibility. To rationalize the pattern of observed

earnings, individuals would need to have a small discount factor, no larger than 0.86. Together

these estimates suggest that either these individuals have extreme values for fexibility, and are

outliers in their preferences, or individuals are perhaps procrastinating searching for a new job and

not fully optimizing. In fact, both may be at play, and policy intervention would be benefcial to

mitigate behavioral issues if they are at play.

This paper relates to a large theoretical and empirical literature on the behavior of the unemployed. In particular, the gig economy provides a new option for individuals during this time.

The key contributions that I make in this paper are quantifying takeup of gig work following an

unemployment shock, evaluating its ability to help buffer the drop in income, and estimating the

long-run consequences of this takeup for earnings, employment, and skills acquisition.7 This has

6 Ameriks et al. (2017) show that among older workers, willingness to work longer is higher for jobs that offer fexible

schedules and demand-side factors (such as the availability of such fexible positions). Additionally, several papers

examine bridge jobs, which individuals use to partially retire (Maestas, 2010; Rubert and Zanella, 2015; Ramnath et

al., 2017).

7 In particular, there are several key mechanisms that research has examined in the context of unemployment that relate

to this paper: duration (Mofftt, 1985; Katz and Meyer, 1990a; Chetty, 2008), search intensity (Mortensen, 1977),

reservation wages, consumption smoothing (Gruber, 1997), spousal labor response (Cullen and Gruber, 2000) and job

3

parallels as well to prior work examining the impacts of experience at temporary help agencies on

subsequent labor market outcomes (Autor, 2001; Autor and Houseman, 2010; Pallais, 2014).

Second, this research builds on an emerging but rapidly growing literature that seeks to understand the recent growth of the gig economy and alternative work arrangements, more generally.8

Relatedly, Katz and Krueger (2017) show with survey data that unemployment is a strong predictor of alternative work transitions. Farrell et al. (2019) document bank account income declining

steadily and involuntary job loss events preceding increases in online platform participation and

revenues. While we now have a grasp on the growth of this type of work, very little is yet known

on the impact of participating in these positions on labor-market outcomes.9 My key contribution

is providing evidence on the causal effects of working in the gig economy following job loss, on

short and long-run labor-market outcomes.

Additionally, this relates closely to many studies that highlight the importance of examining

how the fexibility provided by gig work attracts workers.10 In fact, many drivers cite their preference for fexibility as the reason why they work for Uber (Hall and Krueger, 2018). This paper also

complements Koustas (2018), who shows that rideshare income helps drivers smooth consumption

when facing income fuctuations in their primary job. A key contribution of this paper is to investigate long-term outcomes associated with partaking in gig work, specifcally for those who enter

gig work following an unemployment shock.

This paper proceeds as follows. In Section 2, I provide details on the data and sample construction as well as summary statistics. I then present descriptive evidence on how outcomes evolve

dynamically before and after job loss in Section 3. I explain and motivate my empirical approach

in Section 4. In Section 5, I present my main results for prime-age and older workers. I provide evidence on robustness in Section 5.3. A discussion of these results is included in Section 6. Finally,

I conclude in Section 7.

matches (Acemoglu and Shimer, 1999; Acemoglu, 2001).

8 Early efforts to measure gig work use data from a variety of sources, including: surveys, fnancial institutions, google

trends, and private employers. Estimates indicate that roughly 0.4-1.6 percent of workers are involved in the gig

economy (Harris and Krueger, 2015; Farrell and Greig, 2016; Farrell et al., 2018; Katz and Krueger, 2019), and that

the percentage of workers engaged in alternative work arrangements and contract work is increasing more broadly

(Jackson et al., 2017; Collins et al., 2019; Katz and Krueger, 2019). In earlier work, we show that most of the growth

in self-employment appear primarily to be providing labor services as contractors or freelancers (Jackson et al., 2017)

and is largely being driven the by the growth in gig work mediated through online platforms (Collins et al., 2019).

9 Other studies have examined the effects of Uber or gig availability on the gender gap (Cook et al., 2019b), older

workers (Cook et al., 2019a), auto fnancing (Buchak, 2019), and loan delinquency and credit utilization after being

laid off (Fos et al., 2019).

10 Though many workers are not willing to pay for schedule fexibility, there exists a long tail of workers with high

willingness to pay (Mas and Pallais, 2017). Additionally several papers have examined the value of fexibility as a job

amenity or attribute (Goldin, 2014; Maestas et al., 2018; Wiswall and Zafar, 2018).

4

2 Data, Sample Construction, and Summary Statistics

2.1

Individual Income Tax Returns

I use the universe of individual income tax returns fled in the United States from 2005-2017,

which include both income tax returns that are fled by individuals aggregating all of their earnings

and deductions (e.g. Form 1040, Schedule C), and third-party information returns that are fled

by employers or payers on behalf of the payee denoting the amount of money transferred between

the two entities (e.g. W-2s, 1099s). Taken together, these forms contain a wealth of data on:

demographic and economic characteristics, sources of earnings, beneft coverage, and receipt of

social insurance.

Key advantages of these data are the panel nature, which allows me to track individuals over

time, and the comprehensive scope that identifes both the sources and concentration across sources

of an individual’s earnings. Crucially, I am able to differentiate between traditional wage employment and self-employment, and separately identify employers or payers from whom they received

payments.11 As this is administrative population-level data on the self-employed and gig work

population, this addresses many shortcomings of other data sources, which often focus on particular samples, primary employment, or a snapshot in time. This is of particular importance as recent

research has shown that survey-based measures appear to underestimate self-employment Katz and

Krueger (2019); Abraham et al. (2018).

However, these data are designed for tax administration purposes. Thus, the data only contain

the necessary information for an individual to compute and fle their taxes, and for the government

to monitor tax compliance. For example, the data contain aggregate annual earnings from each

employer, but do not decompose the earnings further into hours or a wage rate.

Geography I use counties as the level of geography in my analyses and this is the level for

which I defne the local labor market. The data identify individual addresses including ZIP code

which I map to counties. Individuals typically fle and/or receive multiple tax forms in a given

tax year, each of which do not necessarily contain the same ZIP code. Thus, for the subset of

individuals for whom I identify multiple ZIP codes in a given year, I use the ZIP code that they

denote on their individual tax return (Form-1040) if they fled their taxes.12 For non-flers, I use

the modal ZIP code denoted on all information returns that they received.

Demographics Individual demographics include age and gender, and are retrieved from ta11 It is worth noting that not all self-employed individuals will have work that is frm facing but for those who do, they

would receive a Form 1099 from that frm. However, for those who do not interact with a frm or payer we would not

expect a Form 1099.

12 In cases where the ZIP code on an individual tax return is incorrect or missing, I use the modal ZIP code from other

tax forms fled by an individual (e.g. Schedule C), if valid. Otherwise, I use the modal ZIP code denoted on all

information returns received by an individual.

5

bles obtained by the IRS from the Social Security Administration (SSA). More precisely, the data

contain each individual’s date of birth. I construct age as the tax year minus birth year.

Household Structure In years that individuals fle their taxes, for married individuals I can

match them with their spouse with their individual tax return (Form-1040).13 Similarly, I identify

the number of children that individuals have based on the number of child dependents they claim

(on Form-1040) in that year.

Social Insurance In addition to measures of income, tax information returns contain details on

social insurance receipt. Precisely, I identify the receipt of unemployment compensation (1099G), Social Security Disability Insurance (1099-SSA), and Social Security Retirement beneft withdrawal (1099-SSA).

2.2

Measuring Gig Work and Availability

I identify individuals who provide services through online gig platforms based on the receipt and/or

flling of a variety of tax forms, and irrespective of the tax fling status of that individual. More

specifcally, I utilize information returns (Form 1099-MISC and Form 1099-K) that are distributed

from platforms to workers, and returns fled by an individual denoting self-employment income

(Schedule C).14

I identify individuals who work for gig platforms by those who receive a 1099-K or 1099-MISC

from a gig platform Employer Identifcation Number (EIN), and based off their self-described

business professions on Schedule C. It is important to note that Forms 1099-MISC and 1099-K are

not used solely for the online platform economy; similarly, individuals fle Schedule C to report

income earned from a plethora of sources. Appendix B discusses these forms and their uses in

more details.

Based off publicly available lists, I include approximately 50 large online gig platforms for

whom workers provide labor based work to be apart of the “gig economy” defnition.15 As noted

in Farrell et al. (2018), the majority of platform work is made up of transportation platforms. For

each labor platforms, I identify all individuals who receive a 1099-MISC and/or 1099-K from that

frm. I consider these individuals to have gig work. Additionally, individuals reporting compensation related to one of these platforms, or gig work more broadly, are also included for reasons

discussed in Appendix B.

13 This is true if their fling status is "married fling jointly" or "married fling separately".

14 Information returns are sent from platforms to the IRS regardless of whether an individual ultimately fles their taxes.

15 This defnition of online gig platform work can be viewed as an underestimate of the gig economy. Nonetheless, it

provides extensive coverage of individuals working for major platforms in the sector, and thus works well in terms of

capturing individuals propensity for gig work. Further, given that fling thresholds appear to not be binding in practice

during this time period, there are not concerns in censoring of the earnings distribution among gig workers.

6

Gig Availability by County

I construct a measure of gig availability at the county-by-year level utilizing the prevalence of gig

work as identifed at the individual level. I aggregate the number of individuals that I observe in

each county-by-year cell with any amount of gig work, and identify the frst year of gig availability

to be the frst year in which I observe at least 30 individuals in that county with gig income.16

Appendix Figure A3 highlights the variation across counties over time in the availability of gig

work as defned by this measure. There are two key takeaways from Appendix Figure A3. First,

each year more counties have gig platforms available as an option, indicated by more polygons

being shaded. Second, gig platforms are also becoming more prevalent within each county over

time, indicated by the shading of each polygon shifting from a light beige towards a darker red.17

Gig platform work may be increasing in prevalence over time, within an area, for many reasons.

For instance, this measure includes multiple platforms and thus once one platforms enter others

likely follow suit. Additionally, supply and demand for platforms will grow with local knowledge

of platforms and their services.

2.3

Sample Construction

Pulling together the components of the data described above, I construct my analysis sample to

consist of individuals experiencing involuntary periods of unemployment as identifed by receipt

of unemployment compensation.18 I draw my sample from the population of individuals with any

positive unemployment compensation in the years 2008-2015. Additionally, I utilize data from

2005-2017 for each individual which provides a sample that is balanced in event time over the

period three years pre- and post-UI receipt (including the frst year of UI receipt).

The fnal analysis sample includes an individual’s frst new UI claim between 2008 and 2015.

UI events are restricted to counties that gig platforms enter by 2015. There are 23 million prime-age

workers and 5 million older workers who experience an UI event in my fnal analysis sample, from

which I draw my stratifed random sample. This leaves 917,128 prime-age workers and 106,243

older workers after stratifcation. See Appendix B.3 for more details on each data restriction and

sampling methodology.

16 I utilize a cutoff of 30 individuals in a county-by-year cell for disclosure reasons at this geographical level.

17 Note that the legend is the same across each sub-fgure to facilitate comparison across years.

18 I only observe eligibility for unemployment compensation conditional on take-up of benefts, and not the full set of

individuals who were eligible but chose to not apply. Given the scope of the tax data, there are limited alternative

ways to identify individuals experiencing a period of involuntary unemployment, and the main alternative would be to

examine mass layoffs. See Appendix C for further discussion of UI takeup.

7

2.4

Variable Defnitions

I construct several key outcome variables that together encapsulate the implications for employment, earnings, education, and access to various benefts. All years I use are tax years which

correspond to calendar years.

Gig Work and Gig Earnings Gig work is an indicator for whether or not I identify an individual

with any gig work in a given year, as I described above. Gig earnings represent gross receipts, as

reported by the platforms on Form 1099-MISC and 1099-K, and do not account for the associated

business expenses that an individual deducts. I observe the amount of business deductions that an

individual claims in a given year (on the return Schedule C), and thus also identify self-employment

income net of business deductions. My measure of income that I describe next accounts for all

business deductions claimed by an individual.

Individual and Household Income I use two different measures of income: one at the individuallevel and another at the household-level. I can only identify household-level income for flers when

I can link individuals together with their spouse based off their fling on the Individual Tax Return

(Form 1040). Household-level income I defne as the household’s adjusted gross income (AGI) as

reported on Form 1040. Since I also observe income for non-flers, I construct a second income

measure to incorporate this additional information.19 For individual income, for flers I assign half

of the household’s adjusted gross income for married individuals and all of AGI for non-married

flers, and for non-flers I aggregate income reported on information returns which include wage

earnings (Form W-2), unemployment benefts (Form 1099-G), and social security and disability

benefts (Form SSA-1099).20 As a robustness, for both measures of income I utilize only the

summed information return values rather than AGI.

Labor Force Participation I examine labor force participation across three different types of

employment: traditional wage employment, gig employment, and self-employment more broadly.

For each employment type, I evaluate both the extensive margin, measured as a dummy variable

indicating any amount of employment of that type, and the intensive margin, measured in earnings.

Unfortunately, I do not have data on hours, months or days worked during the year so I cannot break

these earnings apart into an hourly or monthly rate, and thus can only look at aggregate earnings.

SSDI and Social Security Retirement Benefts I construct a variable that identifes the amount

of social security benefts received in a year, as denoted on Form 1099-SSA. I differentiate between

benefts received from the retirement fund versus disability fund.

Post-Secondary Attendance I identify if an individual is a student at a college, vocational

19 Importantly, the individual income measure should be immune to differential changes in fling behavior from gig

availability or take up, that may be present with the household-level income for which I have to restrict to the subsample

of flers.

20 Prior to 2007, due to data limitations, self-employment income cannot be separately allocated to individuals within a

household and thus this allows for consistency over time in the defnition of individual income.

8

school or other post-secondary institution by receipt of Form 1098-T in a given year. All institutions eligible for the Department of Education’s student aid programs must issue this form to all

students and the IRS, which identifes all qualifed education expenses.

2.5

Summary Statistics

In Table 1, I present summary statistics for the entire analysis sample described in Section 2.3. An

observation is an individual-year, and summary statistics are presented for pre-UI receipt years in

the balanced sample restricted years —the three years prior to UI receipt.

As seen in Table 1, the majority of my sample are flers, 91%. 30% of the sample is female

and on average individuals are 33 years old. Among the 91% who fled, household AGI is on

average $45,997 and the median household income is $34,500.21 Individual income is on average

$31,920 and almost entirely attributable to wages which are on average $33,204.22 Additionally,

the individual is typically the primary wage earner within the household, as spouses’ wages are on

average $8,362 relative to the individual’s wages, $33,204.

On average over the three years prior to unemployment, 91% held a wage job. As seen in Figure

A1 the share with a wage job is increasing over the three years prior to UI receipt. This is high by

construction, as to receive UI the individual must have frst held a wage job from which to become

unemployed and eligible for UI.23 Additionally, 15% of households fled Schedule C for income

earned in a sole proprietorship —this includes individuals who earn income as an independent

contractor or small business owner, for example. 17% were enrolled at a post-secondary institution.

As a baseline, about 0.30% held a gig position in the pre-UI period.

To help put these numbers in context, compared to the overall wage earning population, these

individuals are on average younger, less likely to be married, and have lower household AGI. For

example, the median household income in the US in 2017 was $61,372, and $55,000 (in 2017

$) back in 2010.24 On the other hand, the median household AGI among this sample, $34,500,

is substantially lower. As a result, a slightly higher number of these individuals, 23%, live in

households that claimed the EITC.

21 I winsorize the top and bottom 1% of income values. Large outlying negative values of AGI typically represent large

claimed losses. The top 1% of wage values are also winsorized.

22 There are a number of reasons why on average wages are slightly higher than individual income. First, a component

of individual income is AGI, which accounts for specifc deductions. Second, is if the household has any reported

business losses then that would reduce the overall AGI. Third, is by the construction of the individual income variable

- if the individual is in a married household and earns the majority of the household income from his or her wages,

then when that is divided by the number of two that may be less than his or her wages.

23 Note, an individual may have lost their job at time -1 or 0, and thus we wouldn’t necessarily expect to have 100%

wage job share in either year.

24 Median household income in 2017 was sourced from https://www.census.gov/library/publications/2018/demo/p60263.html.

Median US household income in 2010 was $49,445 and adjusted to 2017 dollars

(https://www.census.gov/newsroom/releases/archives/income_wealth/cb11-157.html).

9

3 Descriptive Evidence on Behavior around UI Receipt

Figures A1 and A2 illustrate how each key outcome variable typically evolves, on average, dynamically relative to the year of UI receipt. These values are restricted to individuals without gig

availability to provide a baseline comparison of magnitudes for subsequent analyses. I describe

outcomes for prime-age workers in Figure A1 and older workers in Figure A2.

Prime-Age Workers

On the extensive margin, measured as indicator for having any gig work in a given year, participation in gig work prior to job loss is effectively zero, which is by construction given that these

individuals at the time of UI receipt live in a county without gig platforms yet available.25 Even

by four years after after UI receipt, only 0.4% of individuals have any gig work compared to 2%

of those who had gig platforms available.

Prior to job loss, household AGI is on an upward trajectory and increases on average from

$40,000 to $49,000. However, households experience a drop of almost $5,000 in the year of UI

receipt and the following year before starting to recover. Similarly, individual income drops from

about $35,000 to $29,000, and begins to recover two years after UI receipt. Underlying these drops

in income are decreases in wage earnings corresponding to the job loss prompting UI receipt.

An individual’s annual wage earnings combine two margins of variation: the extensive margin,

whether an individual holds a wage and salary job, and conditional on having such a position, how

much do they earn in annual earnings. On average, individuals are more likely to hold a traditional

wage and salary position over time prior to the job loss incident of interest. This rate increases

from 87% to 95% of individuals in the years leading up to UI receipt. Recall, by construction,

all of these individuals must have held a wage job at least once in the years prior to the job loss I

identify in order to be eligible for UI. In the year after UI receipt, only 78% of individuals without

gig availability have a wage job, this is a sharp drop of 17 percentage points. Two years following

UI receipt we see an increase of about 5 percentage points in likelihood of holding a wage job,

indicating that roughly one-third of individuals are able to return to the work force. However, this

share stays relatively constant and does not appear to increase substantially over time following

the initial increase.

Corresponding to the job loss timing and the patterns we see with the extensive margin of

holding a wage job, wage earnings increase from about $27,000 to $37,000 prior to UI receipt. In

the year of UI receipt, wage earnings drop to about $27,000 and bottom out at $23,000 the year

25 It is possible that individuals move over time and thus since I have defned gig availability for each individual to be in

the year and county of UI receipt, some individuals may have lived in an area with gig platforms prior to the year of UI

receipt; for this reason, these means are not necessarily precisely zero. However, we see that this is very uncommon,

and individuals appear to have essentially no gig work prior to job loss.

10

following UI receipt before starting to recover on a trajectory similar to that prior to job loss.

The spouse of the individual facing job loss earns about $9,000 prior to the individual’s job

loss and this is relatively stable, though growing slightly, over the years leading up to UI receipt.

Following job loss, we observe a clear spousal labor response given a shift in slope of zero to a positive slope of wage earnings. This indicates that, on average, an individual’s spouse is contributing

more to the household income following the unemployment shock.

On average, about 19% of the individuals are enrolled in post-secondary institutions with eligible tuition payments fve years prior to job loss. This is trending downward leading up to job loss,

as individuals are less likely to still be in school as they age. There is a clear break in trend and

a small increase of about 1 percentage point in schooling in the two years surrounding job loss,

before continuing to decrease on the same trend as prior to job loss.

Older Workers

In Figure A2, there are some key differences in the patterns exhibited among older workers

exhibit as compared to prime-age workers. First, household AGI and individual income are on

average higher at baseline prior to job loss and relatively stable. This is not surprising as these

individuals are older and later in their career. On average, they experience a larger drop in income

and household AGI, $10,000, compared to prime-age individuals whose income dropped by about

$5,000.

Second, there is a substantially larger drop in the share of individuals holding a subsequent

wage job following job loss, almost 30 percentage points. Additionally, unlike the pattern exhibited

by prime-age workers in A1, there is not a small recovery in the rate of individuals holding wage

jobs in the years following UI receipt. Finally, we see an increase in the share of individuals with

the receipt of SSDI benefts that corresponds to the timing of UI receipt. Prior to job loss, the share

with the receipt of SSDI benefts is steady at about 2%. Starting the year of UI receipt, we see

the share with any receipt of SSDI benefts increase from 2% to 10% four years later. There is a

similar pattern with the share of individuals claiming social security retirement benefts; though

this is trending up more prior to job loss and so does not exhibit as pronounced of an increase

following job loss but rather a change in slope.

4 Empirical Approach

The ideal experiment to identify the causal effect of taking up gig work following an unemployment shock would be to randomly assign individuals into and out of gig work at the time of the

unemployment shock. The difference in the outcomes between the two groups would identify

the treatment effect of taking up gig work, as well as the dynamics of these effects. However, in

11

practice there is presumably non-random selection into gig work after an unemployment shock.

For instance, selection might depend on how large or small of an income shock the individual is

faced with, how likely the individual is to get another wage job, which might be a function of their

prior industry or experience, or their ability to recover lost earnings through other responses (e.g.

spousal labor response).

Since the primary objective of this paper is to identify the causal impact of taking up gig work

during spells of unemployment on individual’s outcomes, I need exogenous variation in take up of

gig work to provide a valid counter-factual behavior during unemployment. To address this, my

empirical approach leverages two key sources of variation: the availability of gig platforms and

individual’s propensity for gig work. First, I exploit geographical variation in the availability of

online gig platforms that arises from the rollout of platforms across counties over time. Second,

I introduce within-area variation that permits me to split individuals into two groups: those who

plausibly would even consider gig work, and those unlikely to take up gig work.

These two sources of variation allow me to identify the group of individuals who would have

taken up gig work had it been available to them at the time of unemployment, but happened to

face an unemployment shock in a county prior to the entry of gig platforms. The following two

subsections describe both of these sources of variations in signifcantly more detail.

4.1

Variation in Gig Availability

I leverage variation in the date at which any online gig platform frst enters city to measure the

availability of gig work in a given city in a given year, or on the intensive margin, incorporating the

“intensity” of gig availability based on characteristics such as the number of frms that are present

in a city or how long gig platforms have been present. Crucially, this methodology identifes the

availability of any gig frm rather than relying on the rollout of one specifc gig platform.26 Suppose

that gig platforms frst entered San Francisco in 2010, New York in 2011, and Los Angeles in 2012,

then the thought experiment would be to compare an individual living in San Francisco after 2010

to a similar individual in San Francisco before 2010 as well as to individuals in New York and

Los Angeles where platforms had not yet entered.27 This closely relates to and builds upon several

papers that have exploited variation driven by the launch of specifc gig platforms, most commonly

Uber, across cities(Brazil and Kirk, 2016; Dills and Mulholland, 2017; Berger et al., 2018; Hall

et al., 2018; Koustas, 2018; Buchak, 2019). 28 A simple difference-in-difference exploiting the

rollout variation assumes that the timing of gig frms’ entry into a city is orthogonal to worker

26 Gig availability relies on individuals working for any of the 50 or so large online platforms from publicly available

lists, as mentioned in Section 2.2.

27 This example is for illustrative purposes only.

28 Mishel (2018) estimates that Uber makes up approximately two-thirds of the gig economy.

12

labor supply decisions.29

Figure 2 illustrates the geographical variation in gig availability and prevalence across counties

in 2013 and 2016.30 Each map of the US shows geographical variation by county in the percent

of the working age population, those 15-64, with any earnings from gig work. Darker red shaded

areas indicate a larger percentage of the county working age population have any earnings from

gig work, while lighter beige colors indicate a smaller percentage. Substantial variation in this

measure exists across counties over time. For example in 2013, in the counties surrounding the

Los Angeles area less than 0.14% of working age individuals had any amount of gig work, whereas

in 2016 counties surrounding most major US cities had up to 8.84% of the working age population

with any gig work.

There are two important takeaways from the exhibited variation in the prevalence of gig work.

First, over time more counties have any amount of gig availability as the various platforms rollout

to new geographic markets. Second, the prevalence of gig work continues to increase within an

area following the entry. This is driven by a myriad of factors that include the dissemination of

information regarding a platform’s presence that occurs naturally over time, increased demand for

the services offered by a platform as more consumers become familiar with them, and the entry of

additional platforms as not all necessarily enter a market in the same year.

Approximately 57% of the analysis sample faces an unemployment environment in which gig

platforms were available to them in the year and county in which they received UI. Figure 3 shows

the distribution of the selected UI events across years among those who had gig platforms available

to them at the time of the unemployment shock in the solid gray line with shaded bars, and the

distribution among those who did not have gig platforms available in the dashed black line with

no shading. Not surprisingly, those without gig platforms available are concentrated slightly more

towards the earlier years. As platforms roll out over time, only more individuals will have them

as an option. In Table 2, I show balance on key observable characteristics prior to job loss for

individuals with and without gig availability at UI receipt.

I defne a measure ‘Gig Intensity’ that captures the magnitude of gig availability rather than

a simple indicator for gig work being available as an option. In this measure, I want to capture

any exogenous variation driven by overall trends in how the popularity and availability of these

platforms grow after entering, on average, and exclude variation in the speed of growth arising

from better or worse labor-market outcomes or prospects for workers in that area. Figure A7 plots

the percentage of the working age population in a county-by-year relative to when gig platforms

were frst introduced in that county. A linear approximation appears to roughly ft the average

29 Reiterating Footnote 3, the year of entry of gig platforms in a county is largely predicted by population.

A simple

regression of the year of gig entry on 2014 county population, as a cubic to pick up the curvature of the relationship,

has an R squared value of 0.75.

30 Appendix Figure A3 presents the same map year by year.

13

growth in gig prevalence following the entry of platforms into a county.

I consider the “treatment” of gig availability to occur at the time of UI receipt. Thus, as a

function of the county c in which individual i lives in at the time of UI receipt ti0 , I defne Gig

Intensity as:

Gig Intensityi =

(# Years Gig Available)c(i,t 0 ),t 0

i

i

maxi (# Years Gig Available)c(i,t 0 ),t 0

i

= 19 (# Years Gig Available)c(i,t 0 ),t 0

i

i

i

I rescale this measure to be between 0 and 1 by dividing by 9, the max number of years gig

platforms had been available in the county and year of UI receipt over all individuals. A value

of 1 can be interpreted as becoming unemployed in an environment with the most gig availability

relative to a value of 0 which indicates no gig availability. Among those individuals with any gig

platforms, the median gig intensity value is 13 .

Appendix Figure A6 provides an example of how the gig intensity value varies across areas

and years. Individuals receiving UI in years prior to gig platform entry will have a value of 0.

For example, individuals in County C in Appendix Figure A6 who experience their job loss in the

years 2008-2010 would have a gig intensity value of 0. Gig platforms enter County C in 2012,

those receiving UI in 2012 would therefore have a value of 19 , 2013 would have a value of 29 , and

so on.

4.2

Propensity for Gig Work

Since this time period covers the Great Recession as well as the post recession recovery, and

the timing of entry of frms is correlated with population, it’s plausible that labor-market outcomes in larger cities were recovering at different rates compared to smaller cities. Thus, a simple

difference-in-differences would confound these two effects. Therefore, I employ a third difference

that allows me to incorporate a within-area variation to pick up on local labor market changes.

I utilize pre-UI characteristics to predict an individual’s propensity for gig work. Among treated

individuals, the subset of individuals that had gig platforms available at UI receipt, I observe who

takes up gig work and who does not. Thus, I estimate a probit regression of gig take up on pre-UI

characteristics such as income, wages, EITC claiming, and demographic characteristics. With the

probit estimates I predict a gig propensity for each individual, including those who did not have

access to gig platforms. I split the sample into high-gig-propensity and low-gig-propensity, trying

to capture all the potential gig workers in the high-gig-propensity group and everyone else in the

low-gig-propensity group.

More specifcally, I estimate a probit function where I look at gig take-up post UI as a function

of the following pre-UI characteristics. First, I use demographic characteristics in the year prior

14

to unemployment insurance claiming (i.e. at event time 0). These include a polynomial of an

individual’s age and their gender. I also utilize the zip code in which he or she lived in that

year. Second, I use economic outcomes for the three years leading up to the claiming of UI. This

incorporates a polynomial of wages, income, whether or not an individual had a wage job, any

income from a sole proprietorship, and the share of the household earnings that the individual

contributes, for those fling jointly. Third, I include fling status, marital status and number of

claimed children living in the household.31 Finally, I utilize information about the payer EIN in

the year prior to UI receipt, as this provides additional, otherwise observable, information about the

worker’s characteristics and the likelihood of taking up gig work after losing their job at this frm.32

Though I cannot identify an individual’s exact occupation, I use 3-digit NAICS codes associated

with the frm’s EIN capturing information about the subsector in which that worker used to work.

I split the sample into three groups: high-gig-propensity, the top 1% of the sample in predicted

gig probability; low-gig-propensity, the next 49% of the sample; and those excluded, the bottom

50% of the sample. By dividing the sample, I hope to capture all of the potential gig workers in

the high-gig-propensity group. The low-gig-propensity group will also help to provide additional

within-area variation for identifcation, as I describe in section 4.3. Additionally, I exclude the

bottom 50% in predicted gig probability from the low-gig-propensity individuals to maintain a

group of individuals that are more comparable to the high-gig-propensity individuals. I present the

distribution of gig propensity scores separately by whether or not gig platforms were available in

the county and year when an individual received UI in Appendix Figure A9a. Appendix Figure

A9b shows the distribution of predicted gig propensities among the low group that I retain.

Summary Statistics on High and Low-gig-propensity Individuals

Table 3 presents summary statistics separately for high-gig-propensity and low-gig-propensity

individuals. Relative to the low-gig-propensity sample of UI recipients, high-gig-propensity individuals are slightly more likely to be female (32% versus 30% female), and less likely to be

married (28% versus 29%) or have children (37% vs 40%). On average, high-gig-propensity individuals have a lower household AGI, $40,565 (vs $46,130), and individual income, $27,713 (vs

$32,025). They’re also more likely to have held a gig work position in the pre-UI years, 0.16% vs

0.01%; though both groups have very low baseline values in this regard.

Figure A8 highlights additional variation in individuals’ pre-UI characteristics and how they

relate to the predicted gig propensity measure. For each, I plot the average predicted gig propensity

by binned values for a few of the key predictive variables. Figure A8a demonstrates that younger

31 These measures are conditional on fling, so for non-flers I code these individuals as single and without children, as

done in Yagan (2019).

32 For individuals with multiple W-2s, I use the payer EIN from the W-2 with the largest amount of wages.

15

individuals are more likely to work in the gig economy. This measure peaks around age 25 and,

though not illustrated here, those below 25 have slightly lower propensities on average. Figure

A8b and Figure A8c show that those with lower incomes and wages two years prior to UI receipt

are more likely to participate in gig work. Figure A8d shows that individuals whose wages make

up a larger share of the household’s total wages are more likely to take up gig work, suggesting

that these individuals are more likely to be the primary earner for the household.

4.3

Estimation Approach

I estimate a difference-in-difference-in-differences (DDD) specifcation that leverages variation

in the availability of gig platforms at UI receipt and in individuals predicted propensity for gig

work, as in Gruber (1994). Relative to a standard difference-in-differences, this strategy incorporates additional within-area treatment information. Since the availability of gig platforms should

differentially affect the high-gig-propensity individuals compared to low-gig-propensity individuals, this additional interaction will help control for any other overall changes that coincide with

treatment that affect the outcomes of all individuals, both high and low-gig-propensity. To the

extent that there are other changes that affect all individuals unemployed in an environment with

gig availability relative to no gig availability that are unrelated to the presence of gig platforms,

incorporating low-gig-propensity individuals should account for these changes.

Formally, my estimating equation quantifying work in the gig economy is as follows:

Gigict = αi + β1 Pit + β2 (Pit ∗ Hi ) + β3 (Pit ∗ Gi ) + β4 (Pit ∗ Gi ∗ Hi ) + λct + ηa(i) + ΓXit + εict

(1)

Pit denotes that year t is post UI receipt for individual i, this includes the year of UI receipt. Hi

is an indicator variable that an individual has a high predicted gig propensity. Denote Ei as the frst

year of UI receipt for individual i. Then Gi measures the intensity of gig availability that individual

i faces at event time 0, t = Ei in the county in which they live, ci,t=Ei . I include individual fxed

effects, county-by-year fxed effects (λct ), and single year-of-age fxed effects (ηa(i) ).

With individual fxed effects, the effects are identifed based off within-individual deviations

from their mean outcome value. County-by-year fxed effects allow me to control for local labor

market shocks that affect all individuals. My identifcation is driven by variation across individuals

who do versus do not have gig platforms available to them, accounting for any existing differences

across these counties and years as identifed by differences across these groups among low-gigpropensity individuals. The key identifying assumption is that changes in the difference between

high and low-gig-propensity individuals are not correlated with the intensity of gig availability.

The coeffcient of interest is β4 . Since I have rescaled the gig intensity measure to be between

0 and 1, a value of 1 indicates becoming unemployed in an environment where gig platforms had

16

been available the longest amount of time, among all individuals in my sample. Thus, the interpretation of the coeffcient is the effect for a high-gig-propensity individual who became unemployed

in an environment where gig platforms had been available for the maximum time relative to not being available at all, netting out any differences occurring overall captured by the low-gig-propensity

group. The effect is estimated linearly in the treatment measure of gig availability so scaling the

coeffcient provides the effect size for a given treatment level. So, to get the effect of frst receiving

UI in an environment that had 50% of the maximum gig availability, a county and time combination where gig platforms had been available for half the number of years compared to the longest

available, then you would multiply the coeffcient by one half.

To estimate the effect of working in the gig economy following job loss on labor-market outcomes, I estimate an analogous set of reduced-form regressions with labor-market outcomes, Yict ,

as the dependent variable.

Yict = αi + β1 Pit + β2 (Pit ∗ Hi ) + β3 (Pit ∗ Gi ) + β4 (Pit ∗ Gi ∗ Hi ) + λct + ηa(i) + ΓXit + εict

(2)

Equation 2 estimates reduced-form estimates and identifes the causal effect of gig availability

on unemployment outcomes. Scaling by the frst stage take-up of gig work in Equation 1 would

identify a “treatment on the treated” effect, or the effect of taking up gig work on unemployment

outcomes in this context. This requires stronger assumptions: exclusion of the instrument and

monotonicity (Angrist and Imbens, 1994). Taken together, these assumptions imply that the estimated changes among the high-gig-propensity individuals in earnings, labor force participation,

and schooling, are only due to changes in the those who took up gig work.

5 Effects on Labor Supply and Earnings

I frst present the results graphically with the coeffcient of interest β4 from Equation 2 split into

year by year coeffcients rather than just post. Equation 3 is exactly analogous to Equation 2 above,

but a dynamic version. Event time, in years relative to UI receipt, is denoted by k. In each fgure, I

exclude the year two years prior to UI receipt, and so the coeffcient estimates are relative to event

time −2. Given the structure of the tax data, since I observe unemployment compensation at k = 0

it is possible that job loss occurred in year prior k = −1. Thus, to be conservative, I choose k = −2

to be the excluded year.

 









Yict =αi + ∑ θ1,k Titk + ∑ θ2,k Titk ∗ Hi + ∑ θ3,k Titk ∗ Gi

k=

6 2

k6=2

k6=2





k

+ ∑ θ4,k Tit ∗ Gi ∗ Hi + λct + ηa(i) + ΓXit + εict

k6=2

17

(3)

Titk = 1{t = Ei + k} represents a dummy indicating event time relative to the frst year of UI

receipt for individual i, Ei . The coeffcients of interest in this specifcation are θ4,k .

5.1

Prime-Age Workers

Gig Employment and Earnings

First, I examine extensive margin measures of gig work to quantify to what extent individuals start

working in the gig economy after losing their job. Figure 4a shows an increase of 10.16 percentage

points, among high-gig-propensity individuals, in the year of UI receipt relative to two years prior

for those with the most gig availability relative to no gig availability, netting out any changes

occurring simultaneously among the low-gig-propensity individuals. Among low-gig-propensity

individuals, there is no observed increase in gig work following UI receipt and for all individuals

there is a base of roughly zero gig work in the pre-UI period.

In the year following UI receipt, this extensive margin increase is twice as large, a 19.63 percentage points increase relative to two years prior to UI. This is not surprising if we think individuals are waiting until they exhaust UI benefts before entering, then we would expect a distribution

across months in when individuals exhaust UI benefts. In expectation, only about half of individuals would actually exhaust UI benefts in the same tax year as I frst observe them receiving

unemployment compensation, given a typical state’s UI duration of 26 weeks and assuming UI

recipients are randomly distributed throughout the year. As I only observe the year in which an

individual receives unemployment compensation and not the month, I cannot differentiate between

those starting UI benefts in February versus November.

The next important result from Figure 4a is that gig work increases among high-gig-propensity

individuals following unemployment and then does not decline even four years after UI receipt. If

individuals were using this only for a short period while searching for another job, then we would

expect to see an increase in gig work but then a subsequent decrease when they switched to another

job. However, we can immediately see that individuals enter into these positions and then stay in

them.

In Figure 4b, I present the corresponding results with an intensive margin measure of gig work

—gig earnings (in 2017 dollars).33 The 10.16 percentage points increase in gig work corresponds

to a roughly $588 increase in gig earnings in the year of UI receipt. This implies that each individual working in the gig economy in the year of UI receipt is earning on average $5,800 in that year.

The frst year following UI receipt, high-gig-propensity individuals with the maximum gig availability relative to no gig availability experience an increase in gig earnings of $1,851, implying

annual average gig earnings of roughly $9,400. Implied average annual gig earnings for those with

33 Dollars are adjusted using CPI-U from BLS: https://www.bls.gov/cpi/home.htm.

18

gig work two to four years after UI receipt are roughly $14,000. For perspective, this is roughly

similar to full-time equivalent earnings at federal minimum wage.34

As one of my key objectives in this paper is to disentangle short and long-run labor supply

effects and motivated by the dynamic nature of the effects that I present, I separately estimate

regressions for short- and long-run effects rather than pooling all post years, in Tables 4 and 5,

respectively. In all regressions, I exclude one year prior to UI receipt since it is possible that

job loss occurs in this period and therefore this may be a pre-unemployment period for some

individuals and post-unemployment for others. In all regressions event years, k ∈ [−5, −2] are

considered pre-unemployment years. Short-run estimates present the immediate effect in the year

of UI receipt (k = 0) as the post period of interest, while the long-run estimates consider k ∈ [2, 4]

as the post period of interest. Pooled estimates include the entire post period k ∈ [0, 4].

In Table 4, I show that in the year of UI receipt gig work increases on the extensive margin by

10.51 percentage points and $620.3 in gig earnings, for high-gig-propensity individuals relative to

those without gig platforms available. As seen in Table 5, by two to four years after UI receipt the

increase in gig work is 19.82 percentage points and $2,831 in gig earnings. For completeness, I

also present coeffcients where I pool the short and long-run effects for all outcomes in Appendix

Table A1.

Individual and Household Income

Given the increase in gig work, to what extent does this recover lost income with job loss? Figure

5 highlights the dynamic effects of gig availability at unemployment on individual income. First,

there is a clear short-term smoothing effect. Column 3 of Table 4 shows that in the short run

the income of high-gig-propensity individuals with the maximum gig availability when they lost

their job, relative to that of individuals with no gig platforms available, dropped by $3,118 less.

As illustrated in Figure 5, this advantage fades away rapidly. By the year after UI receipt, those

with and without gig availability at UI receipt have comparable and statistically indistinguishable

changes in income.

Two to four years after UI receipt, the income of high-gig-propensity individuals who had

gig platforms available at UI receipt lags behind comparable individuals who did not have gig

platforms available. In Column 3 of Table 5, I present the coeffcient estimate where I pool all

three long-run post years. The coeffcient -$2,478 indicates that high-gig-propensity individuals

with the maximum gig availability at the time of UI receipt is $2,478 lower two to four years

following UI receipt than comparable individuals who did not have any gig availability.

Appendix Figure A10 exhibits an analogous pattern for household income. Column 4 of Table

5 suggests a larger decrease of about $4,847 when accounting for household income rather than

34 $15,000 is roughly 2,000 hours at the federal minimum wage, $7.25. (https://www.dol.gov/whd/minimumwage.htm)

19

just individual income. Point estimates then drop below zero from one year post UI onwards. Together these results suggest that, among high-gig-propensity individuals, those with gig platforms

available when they lose their job are better able to smooth income in the year of UI receipt. However, two to four years later, their income recovery starts to lag behind. The natural question is

what drives this reversal?

Wage Employment and Earnings

Reversal in income recovery is explained by lower wages earnings, Figure 6a. This is largely a

function of the extensive margin, holding a traditional wage job, as shown in Figure 6b. Examining

the long-run outcomes, Column 5 of Table 5 indicates that annual wage earnings of high-gigpropensity individuals who had the maximum gig availability when they lost their job drop by

$3,951 more than those with no gig platforms available when they lost their job. On the extensive

margin, Column 6 indicates a 4.5 percentage points larger decrease in the probability of holding

a traditional wage job. This is in contrast to the short-run, the year of UI receipt, there is no

differential decrease in holding a wage job or wages in Table 4 Columns 5 or 6.35

Recall in Figure 4a that individuals entered into gig work and stayed in these positions even

a few years later. This does not necessarily preclude the possibility of also holding a traditional

wage job given the fexibility of gig work. However, these individuals are likely working close to

full-time. Though I cannot directly observe hours, using estimates from the literature on typical

hourly wages for Uber drivers —$9.21 (Mishel, 2018) —suggests that these individuals were likely

working close to full-time given that on average they were earning roughly $14,000 annually. On

the one hand, these individuals might be working so intensively on gig platforms because they

have not received a job offer to re-enter a traditional wage position, but are searching. On the other

hand, it is similarly possible that individuals are working so extensively that they do not have time

to search for another job. Finally, it may be that these individuals enter these positions following

job loss, learn they value the fexibility offered by gig work, and choose to stay accepting lower

earnings because they value the fexibility.

5.2

Older Workers

Now I turn to older workers, individuals between the ages of 55 and 69 at the time of UI receipt.

Compared to prime-age workers, older workers are particularly vulnerable in that they have a more

diffcult time of fnding re-employment following job loss and thus typically behave differently

following job loss. Furthermore, they may especially value the fexible nature of gig work as a

35 Note there is a decrease in wages in Table 4 Column 4 that is about one-third of baseline wages in the post period for

all individuals; however, the extensive margin coeffcient is small given that many individuals still receive a W-2 in the

year of UI receipt.

20

bridge to retirement (Ramnath et al., 2017).36

Following my empirical strategy for prime-age workers exactly, I estimate an analogous set

of regressions and fgures using Equations 1, 2, and 3 for the sample of older workers. I present

summary statistics for this older age group in Table 6. Among this sub-population (55-69), approximately 80% are under age 65. The average individual is 56 years old (and the median individual

is 55 years old). 37

Gig Employment and Earnings

Compared to prime-age workers, older workers exhibit a similar increase in gig work following job

loss, Figure 7a. Figure 7a indicates that among the high-gig-propensity individuals who became

unemployed in a county and year with the highest gig availability relative to having no gig availability increased gig work by 15.76 and 31.39 percentage points in the year of and year following

UI receipt, respectively. This corresponds to an increase of $1,248 and $4,430 in gig earnings,

Figure 7b.

The extensive margin increase for high-gig-propensity older workers is almost twice as large

in magnitude as observed for high-gig-propensity prime-age workers. Furthermore, above and beyond the larger extensive margin increase in gig work, the implied gig earnings for each gig worker

on average are also higher. Table 7 and 8 indicate an increase in gig work (and gig earnings) of

16.23 percentage points ($1,233) and 33.09 percentage points ($6,423) in the short-run and longrun, respectively. These estimates imply that each gig worker is earning on average $7,600 in the

short-run and $19,400 in the long-run, and are both approximately 30% larger than we saw for

prime-age workers.

Individual and Household Income

Unlike among prime-age workers, older workers with gig availability do not exhibit the same reversal pattern in income relative to those without gig platforms at the time of job loss (see Appendix

Figure A11). If anything, the point estimates in Appendix Table A2 suggest that individuals with

the maximum gig availability maintain the relative increase of about $5,204 in individual income

and $6,000 in household AGI, though the coeffcients are not statistically signifcant. Again, these

coeffcients are for high-gig-propensity individuals who frst receive UI in an environment with

the maximum gig availability (as a function of county-by-time) relative to comparable individuals

36 While Ramnath et al. (2017) fnd transitions into self-employment as a bridge job between career employment and

retirement less common than expected, this may be due to the fxed costs of entering self-employment, which are

higher than working on a gig platform. Additionally, they examine this in the context of overall workforce transitions

where as I examine individuals facing an unexpected job loss and who are therefore likely not ready to retire.

37 Note that these are averages in the years prior to UI receipt, while the sample is restricted to individuals 55-69 at the

time of UI receipt, these individuals will be a few years younger in the years prior.

21

with no gig availability when they received UI.

Social Security Disability Insurance (SSDI)

Their counterparts, without gig availability, receive SSDI benefts and claim social security retirement benefts rather than returning to the traditional wage workforce, as do the prime-age workers.

Figure 8 highlights a pronounced drop in the receipt of SSDI in the post-UI period. These coeffcients indicate that high-gig-propensity individuals with gig availability receive SSDI benefts at

lower rates than those without gig availability at UI receipt. More specifcally, Table A2 shows this

is a signifcant reduction of 5.9 percentage points in the receipt of SSDI benefts for those with the

most gig availability relative to no gig availability. This suggests that these individuals are on the

margin between working and not. Therefore, this has important fscal implications.

Social Security Retirement

Additionally, the increase in gig participation and earnings through gig platforms postpones withdrawing social security benefts. Figure 9 is noisier and there is not an obvious and large drop

in withdrawing Social Security retirement benefts. However, given the large confdence intervals

due to a small sample, I cannot rule out effects that would be substantial. There is a 1.3 percentage

points reduction in the short run and 4.4 percentage points reduction in the long run (two to four

years after UI) in claiming Social Security retirement benefts among individuals with job loss with

the most gig availability relative to no gig availability. The negative coeffcients indicate that those

with gig availability are less likely to withdraw social security benefts relative to those with no

gig availability. This indicates that gig work is crowding out increases in claiming Social Security

retirement benefts that follow UI receipt.

Since only a sub-group of older workers can actually respond on this margin, those ages 6267, I zoom in on this group for power in Appendix Figure E1 and Appendix Table E1. Among

these individuals, there is a 14 percentage points reduction in Social Security retirement benefts

following UI receipt. This can be fnancially advantageous for two reasons. First, as shown in

Shoven and Slavov (2014), delaying benefts is generally actuarially advantageous. Second, by

working longer, they can only increase their future lifetime benefts by potentially increasing the

value of the highest years in their earnings history or decreasing the number of years with no

earnings that are taken into account when calculating an individual’s beneft.

5.3

Robustness and Placebo Exercises

In this section, I address two potential concerns with my main identifcation strategy. First, I

present two plots showing robustness around my defnition of high and low-gig-propensity. At

22

baseline, I defne high-gig-propensity as the top 1% of the predicted propensity distribution, exclude the lowest 50% of the distribution, and defne low-gig-propensity as the remaining middle

49%. In Appendix Figure A12, I examine two alternative defnitions. First, I present results including all of the bottom 99% of the predicted propensity distribution in the low-gig-propensity group.

Second, maintaining my baseline defnition of low-gig-propensity, I instead alter the defnition of

high-gig-propensity to encompass a broader group of individuals, and include the top 3% rather

than 1% of predicted gig propensities.

As seen in Appendix Figure A12a, incorporating the lowest 50% of predicted propensities in

the low-gig-propensity group, if anything increases the point estimates slightly for gig work. On

the other hand, broadening the defnition of high-gig-propensity dampens the measured effect on

gig work among the high-gig-propensity group. This is not surprising, as increasing the scope of

the high-gig-propensity group means more individuals who are less likely to take up gig work are

included. Appendix Figure A12b shows the corresponding estimates for individual income under

each defnition of high and low. Reassuringly, the patterns of individual income are similarly more

muted for alternative high defnitions with lower estimates for gig work. This provides additional

support that the observed changes in income are driven from those taking up gig work. I have

altered the defnition of high to various other thresholds between 1%-5% and observe qualitatively

similar patterns.

Second, since propensity for gig work closely relates to income and the order of platform entry is correlated with city size, another potential concern might be that higher and lower income

individuals in larger versus smaller areas might have differential recovery in income following unemployment. To address this concern, I run a placebo test where I draw a new random sample

of UI recipients from 2002-2005, prior to the availability of gig platforms. Using the same coeffcients from the probit regression described in Section 4.2, I generate predicted gig propensities

for this placebo sample and similarly split them into high and low-gig-propensity. To simulate gig

availability, I subtract 9 years from the frst year of gig availability by county in order to generate a

placebo gig availability measure that maintains the relative ordering of the platform rollout. I then

calculate the gig intensity measure as a function of these new placebo gig entry dates.

I estimate an analogous regression using Equation 3 for the key outcome variable for primeage workers, individual income. As seen in Appendix Figure A13, there is no clear pattern of

differential changes in income post unemployment. Additionally, there is no longer the inverse

U-shape showing the reversal of relative income as seen in Figure 5. Thus, I fnd this reassuring

that the results I fnd are not driven by differential trends post-unemployment across the different

groups.

23

6 Discussion and Mechanisms

Given that a new choice is being introduced, in this case the gig economy, it is perhaps surprising

that in the long run high-gig-propensity prime-age workers have lower earnings. A natural question

is whether or not this is rational. The following are a few mechanisms that may account for these

results. First, individuals may be learning that they value fexibility and are willing to take an

earnings loss for the fexible amenity. Second, on the other extreme, this could be consistent with

a behavioral story. An individual plans to use the gig economy in the short run to get back on their

feet and search for a new job, however they procrastinate (e.g. O’Donoghue and Rabin (2001)) and

the time they spend working on the gig platforms ultimately crowds out the search effort. Finally,

potential employers might view the time spent working for online gig platforms as undesirable, or

if not disclosed on one’s resume it may look like a longer period of unemployment.38

This study can help distinguish which of these mechanisms may be at play. First, building on

the literature that examines the value of fexibility as a job amenity, I use the regression estimates

differentiating long-term differences in individual income as a way to infer individuals’ value for

fexibility. Precisely, individuals taking up gig work earn approximately 12,502 fewer dollars per

year than their non-gig counterparts in the long-run, or 2-4 years after UI receipt.39 On average,

these individuals earn approximately $32,000 annual in individual income, implying that they are

willing to forgo 39% of their earnings in exchange for fexibility.

This number is very high relative to fndings in earlier literature; for instance, (Mas and Pallais,

2017) fnd estimates of an individual being willing to forfeit up to 20% of earnings, on average, in

the most extreme case where an employer can set their schedule on short notice.40 One consideration is that in prior studies, values of fexibility were considered under different contexts and might

not necessarily refect the same degree of fexibility and hence of valuation. This suggests that

either individuals in this context are either on the extreme end of the distribution in their valuation

for fexible work arrangement, or that there is a behavioral story at play.

Furthermore, we can consider what discount rate would an individual need to have in order to

rationalize the pattern of earnings observed among gig workers, with higher short-term earnings

and lower long-term earnings. These individuals would need to have a discount factor of at most

δ = 0.86 (or an implied interest rate of at least 0.16) to rationalize the choice between the present

38 Adermon and Hensvik (2022) show that in Sweden listing a gig job on one’s resume improves callback rates for

individuals with Swedish sounding names and not Muslim sounding names.

39 Using the estimates from Table 5, 19.82% of high-propensity individuals taking up gig work and earn approximately

2,478 fewer dollars than their non-gig counterparts (-2,478/.1982).

40 Note, that value of fexible work arrangements are also considered in the context of the gig economy, specifcally ride

share, in Chen et al. (2019) and fnd high valuation for fexibility, but in comparison with a non-fexible job in a similar

high wage fuctuation scenario, which differs from the considerations here.

24

discounted value of taking up gig work opposed to not.41 Prior studies suggest this is a low discount

factor (Frederick et al., 2002). 42

Estimates may be consistent with a β −δ type model and time-inconsistent preferences (Strotz,

1956; Ainslie, 1991; Laibson, 1997; O’Donoghue and Rabin, 1999, 2001; Frederick et al., 2002;

Augenblick et al., 2015). Under this model, individuals have a present-bias and may continue to

hold off searching until the next day, and choose to work their gig job today. To the extent this

is occuring suggests the need for a policy intervention to help younger workers exit gig work and

return to traditional work after a short-run recovery. Together these estimates suggest that either

these individuals are have extreme values for fexibility, and are outliers in their preferences, or

individuals are perhaps procrastinating and not fully optimizing. Future work can help shed light

on if either or both of these types of individuals are present.

7 Conclusion

In summary, I document an increase in gig work after job loss (UI receipt). DDD estimates indicate

an increase of 18 percentage points in annual gig-work participation post-UI receipt among the

high-gig-propensity individuals. Correspondingly, the high-gig-propensity individuals with gig

platforms available experience a smaller drop in individual and household income in the year of UI

receipt relative to comparable individuals without gig platform availability. However, this income

smoothing advantage is short-lived, and by two years after UI receipt individual and household

income actually start lagging behind their counterparts without gig availability. This is explained

by a reduction in traditional wage employment and wage earnings.

Crucially, the implications depend on what counterfactual behavior is being crowded out. For

prime-age workers (25-54), it appears to crowd out wage jobs that provide individuals with an upward earnings trajectory, offer important employer-sponsored benefts, and are covered by workplace protections. While for older workers (55-69), the new option of gig work prolongs labor

force participation. In doing so, this reduces receipt of SSDI benefts and postpones claiming

of Social Security retirement benefts. Thus, the effects appear positive in terms of the long-run

implications among older workers.

When examining potential mechanisms at play for the prime-age workers, estimates suggest

one of two things. Either individuals have extreme values of fexibility, as compared to values in

prior studies that examine willingness to pay for fexibility, as well as low discount factors. Alternatively, individuals may be subject to procrastination and obsolescence, and not fully optimizing

41 This compares the present discounted value of high-propensity individuals with gig work available to those without

gig work in years 0-4. A discount factor δ ≤ 0.86 is necessary for discounted earnings to be higher for taking up gig

work.

42 There are experimental studies that have found low discount factors; though this is a non-experimental study.

25

their search behavior. To the extent that this is occurring, suggests the need for a policy intervention to help younger workers search for full employment while in the gig economy. While the

gig positions help individuals cushion job loss in the short-run by providing valuable insurance,

individuals are not shifting back to traditional jobs. Future research can help examine the extent to

which extreme preferences versus behavioral biases are at work, and it is possible that both types

of individuals are present.

Beyond the mechanisms at play, the shift towards gig employment is particularly important because the US systems of beneft coverage and tax administration depend on the employer-employee

relationship. Most Americans receive health insurance coverage, retirement plan coverage, and

related benefts from their employer, in large part because of tax preferences that favor employerprovided coverage. While health insurance coverage is improving for these groups with policies

such as the Affordable Care Act (ACA), gaps in coverage for health and, especially, retirement

benefts remain for this growing group of self-employed (Jackson et al., 2017). Hence, changes

in the employer-employee relationship and shifts toward the non-employee workforce have important consequences for beneft coverage, tax administration, and other labor and tax policy-related

issues. Thus, the implications are more complex than simply measuring changes in income.

References

Abraham, Katharine G, John Haltiwanger, Kristin Sandusky, and James R Spletzer, “Measuring the Gig Economy: Current Knowledge and Open Issues.,” 2018. NBER Working Paper

No. 24950.

Acemoglu, Daron, “Good Jobs versus Bad Jobs,” Journal of Labor Economics, 2001, 19 (1),

1–21.

and Robert Shimer, “Effcient Unemployment Insurance,” Journal of Political Economy, 1999,

107 (5), 893–928.

Adermon, Adrian and Lena Hensvik, “Gig-jobs: stepping stones or dead ends?,” Labour Economics, 2022, 76, 102171.

Ainslie, George, “Derivation of “rational” economic behavior from hyperbolic discount curves,”

American Economic Review, 1991, 81 (2), 334–340.

Ameriks, John, Joseph S. Briggs, Andrew Caplin, Minjoon Lee, Matthew D. Shapiro, and

Christopher Tonetti, “Older Americans Would Work Longer If Jobs Were Flexible,” 2017.

NBER Working Paper No. 24008.

Angrist, Joshua D and Guido W. Imbens, “Identifcation and Estimation of Local Average Treatment Effects,” Econometrica, 1994, 62 (2), 467–475.

26

Augenblick, Ned, Muriel Niederle, and Charles Sprenger, “ Working over Time: Dynamic

Inconsistency in Real Effort Tasks *,” The Quarterly Journal of Economics, 05 2015, 130 (3),

1067–1115.

Autor, David and Susan N Houseman, “Do temporary-help jobs improve labor market outcomes

for low-skilled workers? Evidence from" Work First",” American economic journal: applied

economics, 2010, 2 (3), 96–128.

Autor, David H, “Why do temporary help frms provide free general skills training?,” The Quarterly Journal of Economics, 2001, 116 (4), 1409–1448.

Banerjee, Abhijit V, Arun G Chandrasekhar, Esther Dufo, and Matthew O Jackson,

“Changes in social network structure in response to exposure to formal credit markets,” 2018.

Available at SSRN 3245656.

Berger, Thor, Chinchih Chen, and Carl Benedikt Frey, “Drivers of disruption? Estimating the

Uber effect,” European Economic Review, 2018, 110, 197–210.

Brazil, Noli and David S Kirk, “Uber and metropolitan traffc fatalities in the United States,”

American journal of epidemiology, 2016, 184 (3), 192–198.

Buchak, Greg, “Financing the Gig Economy,” 2019. Job Market Paper.

Chen, M Keith, Peter E Rossi, Judith A Chevalier, and Emily Oehlsen, “The value of fexible

work: Evidence from uber drivers,” Journal of political economy, 2019, 127 (6), 2735–2794.

Chetty, Raj, “Moral Hazard versus Liquidity and Optimal Unemployment Insurance,” Journal of

political Economy, 2008, 116 (2), 173–234.

Collins, Brett, Andrew Garin, Emilie Jackson, Dmitri Koustas, and Mark Payne, “Is Gig

Work Replacing Traditional Employment? Evidence from Two Decades of Tax Returns,” 2019.

SOI Working paper.

Cook, Cody, Rebecca Diamond, and Paul Oyer, “Older Workers and the Gig Economy,” AEA

Papers and Proceedings, May 2019, 109, 372–76.

, , Jonathan Hall, and John List, “The Gender Gap in the Gig Economy: Evidence from a

Million Uber Drivers,” 2019.

Cullen, Julie Berry and Jonathan Gruber, “Does Unemployment Insurance Crowd Out Spousal

Labor Supply,” Journal of Labor Economics, 2000, 18 (3), 546–572.

Dills, AK and SE Mulholland, “Ride-Sharing,” Fatal Crashes, and Crime Available at

https://ssrn. com/abstract, 2017, 2783797.

Farrell, Diana and Fiona Greig, “Paychecks, Paydays, and the Online Platform Economy,” 2016.

JPMorgan Chase & Co. Institute.

, , and Amar Hamoudi, “The online platform economy in 2018: Drivers, workers, sellers,

and lessors,” 2018. JPMorgan Chase & Co. Institute.

27

, , and , “Bridging the Gap: How Families Use the Online Platform Economy to Manage

theiir Cash Flow,” 2019. JPMorgan Chase & Co. Institute.

Fos, Vyacheslav, Naser Hamdi, Ankit Kalda, and Jordan Nickerson, “Gig-Labor: Trading

Safety Nets for Steering Wheels,” 2019. Available at SSRN 3414041.

Frederick, Shane, George Loewenstein, and Ted O’Donoghue, “Time Discounting and Time

Preference: A Critical Review,” Journal of Economic Literature, 2002, 40, 351–401.

Goldin, Claudia, “A grand gender convergence: Its last chapter,” American Economic Review,

2014, 104 (4), 1091–1119.

Gruber, Jonathan, “The Incidence of Mandated Maternity Benefts,” The American economic

review, 1994, pp. 622–641.

, “The Consumption Smoothing Benefts of Unemployment Insurance,” The American Economic Review, 1997, 87 (1), 192–205.

Hall, Jonathan and Alan Krueger, “An Analysis of the Labor Market for Uber’s Driver-Partners

in the United States,” ILR Review, 2018, 71 (3), 705–732.

Hall, Jonathan D, Craig Palsson, and Joseph Price, “Is Uber a substitute or complement for

public transit?,” Journal of Urban Economics, 2018, 108, 36–50.

Harris, Seth and Alan Krueger, “A Proposal for Modernizing Labor Laws for Twenty-FirstCentury Work: The ‘Independent Worker’,” 2015. The Hamilton Project, Discussion Paper

2015-10.

Jackson, Emilie, Adam Looney, and Shanthi Ramnath, “The Rise of Alternative Work Arrangements: Evidence and Implications for Tax Filing and Beneft Coverage,” 2017. Offce of

Tax Analysis Working Paper 114.

Katz, Lawrence and Alan Krueger, “The Rise and Nature of Alternative Work Arrangements in

the United States, 1995-2015,” ILR Review, 2019, 72 (2), 382–416.

Katz, Lawrence F and Alan B Krueger, “The Role of Unemployment in the Rise in Alternative

Work Arrangements,” American Economic Review: Papers Proceedings, 2017, 107 (5), 388–

92.

and Bruce D Meyer, “The impact of the potential duration of unemployment benefts on the

duration of unemployment,” Journal of public economics, 1990, 41 (1), 45–72.

and , “Unemployment insurance, recall expectations, and unemployment outcomes,” The

Quarterly Journal of Economics, 1990, 105 (4), 973–1002.

Koustas, Dmitri, “Consumption Insurance and Multiple Jobs: Evidence from Rideshare Drivers,”

2018. Manuscript.

Laibson, David, “Golden Eggs and Hyperbolic Discounting,” Quarterly Journal of Economics,

1997, 112, 443–477.

28

Maestas, Nicole, “Back to work expectations and realizations of work after retirement,” Journal

of Human Resources, 2010, 45 (3), 718–748.

, Kathleen J Mullen, David Powell, Till Von Wachter, and Jeffrey B Wenger, “The Value

of Working Conditions in the United States and Implications for the Structure of Wages,” 2018.

NBER Working Paper No. 25204.

Mas, Alexandre and Amanda Pallais, “Valuing Alternative Work Arrangements,” American Economic Review, 2017, 107 (12), 3722–3759.

Meyer, Bruce, “Unemployment Insurance and Unemployment Spells,” Econometrica, 1990, 58

(4), 757–782.

Mishel, Lawerence, “Uber and the labor market: Uber drivers’ compensation, wages, and the

scale of Uber and the gig economy.,” 2018.

Mofftt, Robert, “Unemployment Insurance and the Distribution of Unemployment spells,” Journal of Econometrics, 1985, 28 (1), 85–101.

Mortensen, Dale T, “Unemployment Insurance and Job Search Decisions,” ILR Review, 1977, 30

(4), 505–517.

O’Donoghue, Ted and Matthew Rabin, “Doing it Now or Later,” American Economic Review,

1999, 89, 103–124.

and , “Choice and procrastination,” The Quarterly Journal of Economics, 2001, 116 (1),

121–160.

Pallais, Amanda, “Ineffcient hiring in entry-level labor markets,” American Economic Review,

2014, 104 (11), 3565–99.

Ramnath, Shanthi, John Shoven, and Sita Nataraj Slavov, “Pathways to Retirement through

Self-Employment,” 2017. NBER Working Paper No. 23551.

Rubert, Peter and Giulio Zanella, “Revisitng Wage, EArnings, and Hours Profles,” Journal of

Monetary Economics, 2015.

Shoven, John B and Sita Nataraj Slavov, “Does it pay to delay social security?,” Journal of

Pension Economics & Finance, 2014, 13 (2), 121–144.

Strotz, Robert H., “Myopia and Inconsistency in Dynamic Utility Maximization,” Review of Economic Studies, 1956, 23, 165–180.

Wiswall, Matthew and Basit Zafar, “Preference for the Workplace, Investment in Human Capital, and Gender,” The Quarterly Journal of Economics, 2018, 133 (1), 457–507.

Yagan, Danny, “Employment hysteresis from the great recession,” Journal of Political Economy,

2019, 127 (5), 2505–2558.

29

Figures and Tables

Figure 1: Number of Gig Workers by Year

Notes: This fgure presents the yearly number of individuals with any gig work as identifed using the universe of

federal individual income tax returns for the US. This includes counts of the universe of individuals who received

Form 1099-MISC or Form 1099-K from a gig platform, as described in Section 2.2, or fled Schedule C denoting

income from one of these platforms.

30

Figure 2: Percent of a County’s Working Age Population (15-64) Partaking in Gig Work

(a) Variation in 2013

(b) Variation in 2016

Notes: Figure 2a and 2b illustrate the geographic variation in gig platform availability at a given point in time across

counties. Second, they illustrate variation within a county over time in the prevalence of gig work, as measured in

the percent of the counties working age population with any amount of gig earnings in that year. “Insuffcient Data”

means that a cell has fewer than 30 observations with any gig work and are suppressed; predominantly, these consist

of zeros rather then suppressed data points.

31

Figure 3: Distribution of Gig Availability Among UI Recipients Across Years

Notes: ‘Gig Available’ denotes the subset of individuals who had gig platforms available in the county and year in

which they frst receive UI benefts and are shown above with a solid blue line with shaded bars. ‘Gig Unavailable’

denotes the subset of individuals who did not have gig platforms available to them in the county and year in which

they frst receive UI and are shown above with a black dashed line and un-shaded bars. The distribution among each

group sums to 1, and 57% of the sample had gig platforms available at UI receipt.

32

Figure 4: Yearly Coeffcients for Gig Work

(Prime-Age Workers)

(a) Gig Work (x 100)

(b) Gig Earnings (2017 )

Notes: Dependent variable in the top panel is an indicator for participating in gig work, expressed in percentage points

(taking the values 0 or 100). Dependent variable in the bottom panel is Gig Earnings (in 2017 $). Coeffcient estimates

for Titk ∗ Gi ∗ Hi are plotted for each year in event time relative to event time k=-2. These estimates are interpreted as the

effect of having the most gig availability at UI receipt compared to no gig availability among high-gig-propensity individuals, differencing out any changes that occur among the low-gig-propensity individuals, and each yearly coeffcient

is relative to two years prior to UI receipt. Sample balanced on event time k ∈ [−3, 2]. Regression includes individual

FEs, county-by-year FEs, and age FEs. Data drawn around UI recipients frst UI event in the period 2008-2015 and

includes observations from 2005-2017. Robust standard errors clustered by individual.

33

Figure 5: Yearly Coeffcients for Individual Income

(Prime-Age Workers)

Notes: Dependent variable is individual income (2017 $), and the top and bottom 1% of values are winsorized.

Coeffcient estimates for Titk ∗ Gi ∗ Hi are plotted for each year in event time relative to event time k=-2. These estimates

are interpreted as the effect of having the most gig availability at UI receipt compared to no gig availability among

high-gig-propensity individuals, differencing out any changes that occur among the low-gig-propensity individuals,

and each yearly coeffcient is relative to two years prior to UI receipt. Sample balanced on event time k ∈ [−3, 2].

Regression includes individual FEs, county-by-year FEs, and age FEs. Data drawn around UI recipients frst UI event

in the period 2008-2015 and includes observations from 2005-2017. Robust standard errors clustered by individual.

34

Figure 6: Yearly Coeffcients for Wage Employment

(Prime-Age Workers)

(a) Wage Earnings (2017 $)

(b) Wage Job (x 100)

Notes: Dependent variable in the top panel is wage earnings (2017 $), and the top 1% of values are winsorized. Dependent variable in the bottom panel is an indicator for wage job (in percentage points 0 or 100). Coeffcient estimates

for Titk ∗ Gi ∗ Hi are plotted for each year in event time relative to event time k=-2. These estimates are interpreted as the

effect of having the most gig availability at UI receipt compared to no gig availability among high-gig-propensity individuals, differencing out any changes that occur among the low-gig-propensity individuals, and each yearly coeffcient

is relative to two years prior to UI receipt. Sample balanced on event time k ∈ [−3, 2]. Regression includes individual

FEs, county-by-year FEs, and age FEs. Data drawn around UI recipients frst UI event in the period 2008-2015 and

includes observations from 2005-2017. Robust standard errors clustered by individual.

35

Figure 7: Yearly Coeffcients for Gig Work

(Older Workers)

(a) Gig Work (x 100)

(b) Gig Earnings (2017 )

Notes: Dependent variable in the top panel is an indicator for participating in gig work, expressed in percentage points

(taking the values 0 or 100). Dependent variable in the bottom panel is Gig Earnings (in 2017 $). Restricted to the

older worker sample, ages 55-69 at the time of UI receipt. Coeffcient estimates for Titk ∗ Gi ∗ Hi are plotted for each

year in event time relative to event time k=-2. These estimates are interpreted as the effect of having the most gig

availability at UI receipt compared to no gig availability among high-gig-propensity individuals, differencing out any

changes that occur among the low-gig-propensity individuals, and each yearly coeffcient is relative to two years prior

to UI receipt. Sample balanced on event time k ∈ [−3, 2]. Regression includes individual FEs, county-by-year FEs,

and age FEs. Data drawn around UI recipients frst UI event in the period 2008-2015 and includes observations from

2005-2017. Robust standard errors clustered by individual.

36

Figure 8: Yearly Coeffcients for Social Security Disability Insurance

(Older Workers)

Notes: Dependent variable is an indicator having received SSDI benefts (in percentage points 0 or 100). Restricted

to the older worker sample, ages 55-69 at the time of UI receipt. Coeffcient estimates for Titk ∗ Gi ∗ Hi are plotted for

each year in event time relative to event time k=-2. These estimates are interpreted as the effect of having the most gig

availability at UI receipt compared to no gig availability among high-gig-propensity individuals, differencing out any

changes that occur among the low-gig-propensity individuals, and each yearly coeffcient is relative to two years prior

to UI receipt. Sample balanced on event time k ∈ [−3, 2]. Regression includes individual FEs, county-by-year FEs,

and age FEs. Data drawn around UI recipients frst UI event in the period 2008-2015 and includes observations from

2005-2017. Robust standard errors clustered by individual.

Figure 9: Yearly Coeffcients for Social Security Retirement Benefts

(Older Workers)

Notes: Dependent variable is an indicator having claimed Social Security Retirement benefts (in percentage points 0 or

100). Restricted to the older worker sample, ages 55-69 at the time of UI receipt. Coeffcient estimates for Titk ∗ Gi ∗ Hi

are plotted for each year in event time relative to event time k=-2. These estimates are interpreted as the effect of

having the most gig availability at UI receipt compared to no gig availability among high-gig-propensity individuals,

differencing out any changes that occur among the low-gig-propensity individuals, and each yearly coeffcient is

relative to two years prior to UI receipt. Sample balanced on event time k ∈ [−3, 2]. Regression includes individual

FEs, county-by-year FEs, and age FEs. Data drawn around UI recipients frst UI event in the period 2008-2015 and

includes observations from 2005-2017. Robust standard errors clustered by individual.

37

Table 1: Pre-UI Summary Statistics

Female

Age

Married

Any Children

Household Filed

Household AGI (2017 $)

Individual Income (2017 $)

Wage Job

Wages (2017 $)

Spouse’s Wages (2017 $)

Schedule C Proft/Loss (2017 $)

Schedule C (HH)

Gig Year

Student

Has SSDI

Has SS Income

Claimed EITC (HH)

Mean Std Dev

0.30

0.46

33

33

0.29

0.45

0.39

0.49

0.91

0.29

45,997 39,667

31,920 25,561

0.91

0.29

33,204 29,286

8,362 19,998

475

3,908

0.15

0.36

0.00

0.01

0.17

0.38

0.0040 0.0627

0.0012 0.0342

0.28

0.45

P10

Median

P90

23

31

45

9,700

4,000

34,500

27,000

98,100

63,900

400

0

0

27,900

0

0

70,200

36,000

0

Notes: Summary statistics are for the three years prior to UI receipt. P10 and P90 represent the 10th and 90th percentile

values of the corresponding variables. All P10, Median, and P90 values are rounded for confdentiality of taxpayer

data. Married is taken from an individual’s fling status. Any Children is an indicator for if a household claimed any

dependents in that year. Household Filed denotes that an individual or their spouse fled an Individual Tax Return

in that tax year (Form-1040). Household AGI is the adjusted gross income drawn directly from the Individual Tax

Return. Individual Income additionally incorporates non-fling earnings by summing information returns (e.g. W-2s,

1099s, etc.) and divides AGI in half for married fling jointly individuals. Wage job is an indicator for receiving a

W-2. Wages is the sum of wages across all W-2 forms received by an individual. Spouse’s Wages is the sum of wages

across all W-2 forms received by an individual; this value is 0 if a spouse has no wages or an individual does not have

a spouse, and is missing for non-flers. Schedule C Proft/Loss denotes the amount of proft/loss claimed on Sch C.

Schedule C (HH) is an indicator for either the individual or spouse having fled Schedule C for income earned through

a sole-proprietorship. Gig Job is an indicator for having an earnings from gig work. Student is an indicator for having

an eligible tuition payment at a post-secondary institution. Has SSDI is an indicator for receiving Social Security

Disability Insurance (Form 1099-SSA). Has SS Income is an indicator for withdrawing Social Security Retirement

Income (Form 1099-SSA). Claimed EITC (HH) is an indicator that a Household claimed the EITC in that tax year.

38

Table 2: Balance Table by Gig Availability

Age

Female

1 Year Prior to UI Receipt:

Wages (’000 $)

Income (’000 $)

Married

Student

Wage Job

HH Wage Share

Tax Filer

Claimed EITC (HH)

Filed Sch C (HH)

2 Years Prior to UI Receipt:

Wages (’000 $)

Income (’000 $)

Married

Student

Wage Job

HH Wage Share

Tax Filer

Claimed EITC (HH)

Filed Sch C (HH)

3 Years Prior to UI Receipt:

Wages (’000 $)

Income (’000 $)

Married

Student

Wage Job

HH Wage Share

Tax Filer

Claimed EITC (HH)

Filed Sch C (HH)

Observations

F-test of joint signifcance

(1)

Gig Unavailable

33.97

0.252

(2)

Gig Available

34.78

0.337

(1) - (2)

P-Value

(0.429)

(0.510)

38.02

47.78

0.346

0.136

0.953

0.888

0.943

0.281

0.163

39.24

48.19

0.286

0.136

0.964

0.910

0.939

0.290

0.153

(0.408)

(0.203)

(0.000)***

(0.027)**

(0.521)

(0.010)**

(0.756)

(0.767)

(0.331)

35.90

45.64

0.339

0.150

0.934

0.884

0.916

0.263

0.162

36.30

45.34

0.279

0.158

0.922

0.906

0.908

0.273

0.156

(0.412)

(0.294)

(0.001)***

(0.014)**

(0.453)

(0.049)**

(0.100)

(0.251)

(0.291)

33.16

43.15

0.327

0.162

0.909

0.883

0.890

0.248

0.163

235,598

33.58

42.55

0.265

0.178

0.885

0.906

0.879

0.255

0.155

499,337

(0.626)

(0.361)

(0.015)**

(0.002)***

(0.364)

(0.156)

(0.965)

(0.288)

(0.070)*

734,935

1.8008

Notes: Columns 1 and 2 present mean values for individuals with UI receipt when gig platforms were and were not

available, respectively. The third column shows the p-values for the difference in sample means controlling for year

FEs and county FEs. Income is individual income as defne in Table 1. Married is taken from an individual’s fling

status. Student is an indicator for having an eligible tuition payment at a post-secondary institution. Wage job is

an indicator for receiving a W-2. HH Wage Share is an individual’s wage earnings as a fraction of the sum of the

individual’s and spouse’s wages. Tax Filer denotes that an individual or their spouse fled an Individual Tax Return in

that tax year (Form-1040). Claimed EITC (HH) is an indicator that a Household claimed the EITC in that tax year.

Filed Schedule C (HH) is an indicator for either the individual or spouse having fled Schedule C for income earned

through a sole-proprietorship.

39

Table 3: Pre-UI Summary Statistics —by Gig Propensity

(Prime-Age Workers)

40

Female

Age

Married

Any Children

Household Filed

Household AGI (2017 $)

Individual Income (2017 $)

Wage Job

Wages (2017 $)

Spouse’s Wages (2017 $)

Schedule C Proft/Loss (2017 $)

Schedule C (HH)

Gig Year

Student

Has SSDI

Has SS Income

Claimed EITC (HH)

Mean

0.32

33

0.28

0.37

0.90

40,565

27,713

0.89

27,459

8,594

600

0.19

0.00

0.19

0.0029

0.0009

0.30

High Gig Propensity

Std Dev P10 Median

0.47

8

23

31

0.45

0.48

0.30

37,278 8,200 29,400

23,386 2,300 22,900

0.32

25,287

0

22,500

21,006

0

0

4,324

0

0

0.40

0.04

0.39

0.0533

0.0298

0.46

P90

45

88,300

56,500

59,200

36,700

1300

Mean

0.30

33

0.29

0.40

0.91

46,130

32,025

0.91

33,347

8,357

472

0.15

0.00

0.17

0.0040

0.0012

0.28

Low Gig Propensity

Std Dev P10 Median

0.46

8

23

31

0.45

0.49

0.29

39,714 9,700 34,700

25,604 4,000 27,200

0.29

29,364

400

28,000

19,972

0

0

3,898

0

0

0.36

0.01

0.38

0.0629

0.0343

0.45

P90

45

98,300

64,100

70,500

36,000

0

Notes: Summary statistics are for the three years prior to UI receipt. P10 and P90 represent the 10th and 90th percentile values of the corresponding variables.

All P10, Median, and P90 values are rounded for confdentiality of taxpayer data. Married is taken from an individual’s fling status. Any Children is an indicator

for if a household claimed any dependents in that year. Household Filed denotes that an individual or their spouse fled an Individual Tax Return in that tax year

(Form-1040). Household AGI is the adjusted gross income drawn directly from the Individual Tax Return. Individual Income additionally incorporates non-fling

earnings by summing information returns (e.g. W-2s, 1099s, etc.) and divides AGI in half for married fling jointly individuals. Wage job is an indicator for

receiving a W-2. Wages is the sum of wages across all W-2 forms received by an individual. Spouse’s Wages is the sum of wages across all W-2 forms received

by an individual; this value is 0 if a spouse has no wages or an individual does not have a spouse, and is missing for non-flers. Schedule C Proft/Loss denotes the

amount of proft/loss claimed on Sch C. Schedule C (HH) is an indicator for either the individual or spouse having fled Schedule C for income earned through a

sole-proprietorship. Gig Job is an indicator for having an earnings from gig work. Student is an indicator for having an eligible tuition payment at a post-secondary

institution. Has SSDI is an indicator for receiving Social Security Disability Insurance (Form 1099-SSA). Has SS Income is an indicator for withdrawing Social

Security Retirement Income (Form 1099-SSA). Claimed EITC (HH) is an indicator that a Household claimed the EITC in that tax year.

Table 4: Short-Run Effects on Labor Supply, Income, and Social Insurance Receipt

(Prime-Age Workers)

VARIABLES

(1)

Gig Year

(x100)

(2)

Gig Earnings

(3)

Individual

Income

(4)

HH AGI

(5)

Wages

(6)

Wage Job

(x100)

Short Run (First Post Year)

0.00469*

(0.00271)

-0.214***

(0.0216)

-0.727***

(0.0664)

10.51***

(0.444)

-0.461**

(0.220)

-7.021***

(2.004)

-46.19***

(6.793)

620.3***

(48.90)

-3,964***

(110.9)

-994.7**

(451.6)

-1,394***

(405.1)

3,118***

(1,039)

-4,932***

(169.5)

208.4

(688.5)

-2,604***

(634.8)

2,573

(1,616)

-11,214***

(132.5)

-770.6

(533.2)

-237.8

(490.6)

1,835

(1,211)

-1.204***

(0.181)

-2.685***

(0.748)

0.236

(0.730)

1.450

(1.847)

Observations (Unweighted)

Observations (Weighted)

R-squared

3,031,317

46,850,202

0.251

3,031,317

46,850,202

0.232

3,031,317

46,850,202

0.811

2,750,404

42,716,803

0.845

3,031,317

46,850,202

0.804

3,031,317

46,850,202

0.442

Pre-Period Dep Var Mean

Pre-Period Dep Var SD

0.01

0.77

0.32

93.24

32,036

25,414

45,178

39,679

33,097

29,275

92.8

25.9

Post

Post x Gig Intensity

Post x High

41

Post x Gig Intensity x High

Notes: Results presented are for the subsample of prime-age workers, those ages 25-54 at the time of UI receipt. Post UI, k ≥ 0, indicate years following (and

including) UI receipt. In these short-run specifcations, Post is restricted to the year of UI receipt. Gig Intensity is a measure, between 0 and 1, of the degree of

gig availability in the county in which an individual lives in at the time of unemployment insurance receipt and is fxed within individual. High is an indicator

variable denoting that an individual is in the high predicted gig propensity sample. Regression includes individual fxed effects, county-by-year FEs, and age FEs.

Data drawn around UI recipients frst UI event in the period 2008-2015 and includes observations from tax years 2005-2017. Top 1% of wages and the top and

bottom 1% of income and AGI are winsorized. All dollar values are infation adjusted to 2017 dollars using CPI-U. Robust standard errors clustered by individual

in parentheses (∗∗∗ p < 0.01,∗∗ p < 0.05,∗ p < 0.1).

Table 5: Long-Run Effects on Labor Supply, Income, and Social Insurance Receipt

(Prime-Age Workers)

VARIABLES

(1)

Gig Year

(x100)

(2)

Gig Earnings

(3)

Individual

Income

(4)

HH AGI

(5)

Wages

(6)

Wage Job

(x100)

Long Run (Two-Four Years Post)

0.00878

(0.0116)

0.0647

(0.0580)

2.090***

(0.140)

19.82***

(0.873)

-2.163

(1.808)

-22.83***

(7.774)

112.9***

(21.88)

2,831***

(167.5)

-7,705***

(165.6)

687.4

(493.0)

432.6

(483.9)

-2,478*

(1,352)

-8,025***

(263.0)

-109.8

(801.5)

-630.2

(814.5)

-4,847**

(2,306)

-13,963***

(194.1)

1,604***

(564.3)

774.6

(589.0)

-3,951***

(1,523)

-14.78***

(0.276)

5.939***

(0.715)

-0.795

(0.829)

-4.542**

(2.252)

Observations (Unweighted)

Observations (Weighted)

R-squared

4,193,859

65,184,888

0.274

4,193,859

65,184,888

0.265

4,193,859

65,184,888

0.721

3,739,463

58,168,178

0.773

4,193,859

65,184,888

0.708

4,193,859

65,184,888

0.406

Pre-Period Dep Var Mean

Pre-Period Dep Var SD

0.01

0.77

0.32

93.24

32,036

25,414

45,178

39,679

33,097

29,275

92.8

25.9

Post

Post x Gig Intensity

Post x High

42

Post x Gig Intensity x High

Notes: Results presented are for the subsample of prime-age workers, those ages 25-54 at the time of UI receipt. Post UI, k ≥ 0, indicate years following (and

including) UI receipt. In these long-run specifcations, Post is restricted to the long-run post years, two to four years after UI receipt. Gig Intensity is a measure,

between 0 and 1, of the degree of gig availability in the county in which an individual lives in at the time of unemployment insurance receipt and is fxed within

individual. High is an indicator variable denoting that an individual is in the high predicted gig propensity sample. Regression includes individual fxed effects,

county-by-year FEs, and age FEs. Data drawn around UI recipients frst UI event in the period 2008-2015 and includes observations from tax years 2005-2017. Top

1% of wages and the top and bottom 1% of income and AGI are winsorized. All dollar values are infation adjusted to 2017 dollars using CPI-U. Robust standard

errors clustered by individual in parentheses (∗∗∗ p < 0.01,∗∗ p < 0.05,∗ p < 0.1).

Table 6: Pre-UI Summary Statistics

(Older Workers)

43

Female

Age

Married

Any Children

Household Filed

Household AGI (2017 $)

Individual Income (2017 $)

Wage Job

Wages (2017 $)

Spouse’s Wages (2017 $)

Schedule C Proft/Loss (2017 $)

Schedule C (HH)

Gig Year

Student

Has SSDI

Has SS Income

Claimed EITC (HH)

Mean Std Dev

0.17

0.37

56

4

0.53

0.50

0.30

0.46

0.94

0.23

65,928 48,402

41,461 29,090

0.90

0.30

43,706 34,507

15,029 25,376

714

5,591

0.28

0.45

0.00

0.01

0.04

0.20

0.01

0.10

0.02

0.14

0.14

0.35

P10

Median

P90

52

55

61

16,500

9,400

53,500

36,300

133,300

79,600

200

0

0

37,700

0

0

89,400

53,500

3300

Notes: Sample is restricted to the older workers subsample, those ages 55-69 at the time of UI receipt. Summary statistics are for the three years prior to UI receipt.

P10 and P90 represent the 10th and 90th percentile values of the corresponding variables. All P10, Median, and P90 values are rounded for confdentiality of

taxpayer data. Married is taken from an individual’s fling status. Any Children is an indicator for if a household claimed any dependents in that year. Household

Filed denotes that an individual or their spouse fled an Individual Tax Return in that tax year (Form-1040). Household AGI is the adjusted gross income drawn

directly from the Individual Tax Return. Individual Income additionally incorporates non-fling earnings by summing information returns (e.g. W-2s, 1099s, etc.)

and divides AGI in half for married fling jointly individuals. Wage job is an indicator for receiving a W-2. Wages is the sum of wages across all W-2 forms received

by an individual. Spouse’s Wages is the sum of wages across all W-2 forms received by an individual; this value is 0 if a spouse has no wages or an individual

does not have a spouse, and is missing for non-flers. Schedule C Proft/Loss denotes the amount of proft/loss claimed on Sch C. Schedule C (HH) is an indicator

for either the individual or spouse having fled Schedule C for income earned through a sole-proprietorship. Gig Job is an indicator for having any earnings from

gig work. Student is an indicator for having an eligible tuition payment at a post-secondary institution. Has SSDI is an indicator for receiving Social Security

Disability Insurance (Form 1099-SSA). Has SS Income is an indicator for withdrawing Social Security Retirement Income (Form 1099-SSA). Claimed EITC (HH)

is an indicator that a Household claimed the EITC in that tax year.

Table 7: Short-Run Effects on Labor Supply, Income, and Social Insurance Receipt

(Older Workers)

VARIABLES

(1)

Gig Year

(x100)

(2)

Gig Earnings

(3)

Individual

Income

(4)

HH AGI

(5)

Wage Job

(x100)

(6)

Has SSDI

(x100)

(7)

Has Soc Sec Ret

(x100)

Short Run (First Post Year)

0.0588***

(0.0155)

-0.375***

(0.0811)

-1.434***

(0.346)

16.23***

(2.205)

3.601**

(1.575)

-25.79***

(7.832)

-112.3***

(34.13)

1,233***

(205.6)

-3,884***

(570.1)

659.8

(1,856)

-2,475

(1,883)

4,174

(4,675)

-4,493***

(896.1)

2,325

(2,960)

-3,425

(2,635)

4,242

(7,776)

-15,024***

(730.2)

-1,280

(2,499)

-111.3

(2,654)

5,778

(7,595)

1.162***

(0.277)

-0.0532

(0.938)

0.640

(1.191)

-2.570

(2.495)

1.540***

(0.468)

1.333

(1.757)

2.774

(2.142)

-1.285

(5.406)

Observations (Unweighted)

Observations (Weighted)

R-squared

165,746

2,697,869

0.303

165,746

2,697,869

0.381

165,746

2,697,869

0.852

154,771

2,547,007

0.880

165,746

2,697,869

0.835

165,746

2,697,869

0.809

165,746

2,697,869

0.746

Pre-Period Dep Var Mean

Pre-Period Dep Var SD

0.01

0.82

0.48

175.88

42,397

28,916

66,580

48,532

44,222

34,283

1.09

10.38

3.20

17.60

Post

Post x Gig Intensity

Post x High

44

Post x Gig Intensity x High

Notes: Sample is restricted to the older workers subsample, those ages 55-69 at the time of UI receipt. Post UI, k ≥ 0, indicate years following (and including) UI

receipt. In these short-run specifcations, Post is restricted to the year of UI receipt. Gig Intensity is a measure, between 0 and 1, of the degree of gig availability

in the county in which an individual lives in at the time of unemployment insurance receipt and is fxed within individual. High is an indicator variable denoting

that an individual is in the high predicted gig propensity sample. Regression includes individual fxed effects, county-by-year FEs, and age FEs. Data drawn

around UI recipients frst UI event in the period 2008-2015 and includes observations from tax years 2005-2017. Top 1% of wages and the top and bottom 1% of

income and AGI are winsorized. All dollar values are infation adjusted to 2017 dollars using CPI-U. Robust standard errors clustered by individual in parentheses

(∗∗∗ p < 0.01,∗∗ p < 0.05,∗ p < 0.1).

Table 8: Long-Run Effects on Labor Supply, Income, and Social Insurance Receipt

(Older Workers)

VARIABLES

(1)

Gig Year

(x100)

(2)

Gig Earnings

(3)

Individual

Income

(4)

HH AGI

(5)

Wage Job

(x100)

(6)

Has SSDI

(x100)

(7)

Has Soc Sec Ret

(x100)

Long Run (Two-Four Years Post)

0.200***

(0.0649)

0.452

(0.285)

1.150*

(0.603)

33.09***

(4.017)

37.00***

(12.91)

51.76

(56.74)

-170.4

(108.0)

6,423***

(768.3)

-10,722***

(852.0)

-1,871

(2,084)

-2,681

(2,030)

6,757

(5,725)

-10,958***

(1,351)

-9,020***

(3,394)

-4,474

(3,185)

10,598

(10,830)

-24,230*** 6.250***

(0.699)

(1,092)

-8.834***

-1,986

(2.095)

(2,586)

-0.759

243.4

(1.624)

(2,716)

-6.227*

11,647

(3.316)

(8,832)

6.238***

(0.868)

6.556***

(2.429)

0.565

(2.427)

-4.390

(6.277)

Observations (Unweighted)

Observations (Weighted)

R-squared

228,361

3,769,690

0.328

228,361

3,769,690

0.326

228,361

3,769,690

0.784

205,439

3,386,606

0.828

228,361

3,769,690

0.752

228,361

3,769,690

0.632

228,361

3,769,690

0.819

Pre-Period Dep Var Mean

Pre-Period Dep Var SD

0.01

0.82

0.48

175.88

42,397

28,916

66,580

48,532

44,222

34,283

1.09

10.38

3.20

17.60

Post

Post x Gig Intensity

Post x High

45

Post x Gig Intensity x High

Notes: Sample is restricted to the older workers subsample, those ages 55-69 at the time of UI receipt. Post UI, k ≥ 0, indicate years following UI receipt. In these

long-run specifcations, Post is restricted to the long-run post years, two to four years after UI receipt. Gig Intensity is a measure, between 0 and 1, of the degree

of gig availability in the county in which an individual lives in at the time of unemployment insurance receipt and is fxed within individual. High is an indicator

variable denoting that an individual is in the high predicted gig propensity sample. Regression includes individual fxed effects, county-by-year FEs, and age FEs.

Data drawn around UI recipients frst UI event in the period 2008-2015 and includes observations from tax years 2005-2017. Top 1% of wages and the top and

bottom 1% of income and AGI are winsorized. All dollar values are infation adjusted to 2017 dollars using CPI-U. Robust standard errors clustered by individual

in parentheses (∗∗∗ p < 0.01,∗∗ p < 0.05,∗ p < 0.1).

FOR ONLINE PUBLICATION:

Availability of the Gig Economy and Long Run Labor Supply Effects for the Unemployed

(Emilie Jackson)

Appendix A

Figures and Tables

46

Figure A1: Summary of Key Outcome Variables by Year Relative to UI Receipt

(Prime-Age Workers)

Gig Job (x 100)

Gig Earnings

HH AGI

Individual Income

Wage Job (x 100)

Wage Earnings

Spouse’s Wage Earnings

Student (x 100)

47

Notes: Averages of each outcome are plotted by year relative to UI receipt for individuals who become unemployed in a county and year where there are no gig

platforms available. Gig Job denotes having any income from gig work in that tax year. Gig Earnings are the sum of all earnings earned in the gig economy.

Household AGI is the adjusted gross income drawn directly from the Individual Tax Return. Individual Income additionally incorporates non-fling earnings by

summing information returns (e.g. W-2s, 1099s, etc.) and divides AGI in half for married fling jointly individuals. Wage job indicates receiving a W-2 in a given

year. Wage Earnings are the sum of all W-2 wages in a given year. Spouse’s Wage Earnings are the sum of all W-2 wages of a spouse and are restricted to flers.

Student is an indicator for having an eligible tuition payment made for post-secondary schooling (Form 1098-T).

Figure A2: Summary of Key Outcome Variables by Year Relative to UI Receipt

(Older Workers)

Gig Job (x 100)

Gig Earnings

HH AGI

Individual Income

Wage Job (x 100)

Spouse’s Wage Earnings

Has SSDI (x 100)

Has SS Ret (x 100)

48

Notes: Averages of each outcome are plotted by year relative to UI receipt for individuals who become unemployed in a county and year where there are no gig

platforms available. Gig Job denotes having any income from gig work in that tax year. Gig Earnings are the sum of all earnings earned in the gig economy.

Household AGI is the adjusted gross income drawn directly from the Individual Tax Return. Individual Income additionally incorporates non-fling earnings by

summing information returns (e.g. W-2s, 1099s, etc.) and divides AGI in half for married fling jointly individuals. Wage job indicates receiving a W-2 in a given

year. Spouse’s Wage Earnings are the sum of all W-2 wages of a spouse and are restricted to flers. Has SSDI is an indicator for receiving Social Security Disability

Insurance (Form 1099-SSA). Has SS Ret Income is an indicator for claiming Social Security Retirement Income (Form 1099-SSA).

Figure A3: Percent of a County’s Working Age Population (15-64) with any Gig Work by Year

2010

2011

2012

2013

2014

2015

Notes: The above maps illustrate the geographic variation in gig platform availability at a given point in time across

counties. Second, they illustrate variation within a county over time in the prevalence of gig work, as measured in

the percent of the counties working age population with any amount of gig earnings in that year. “Insuffcient Data”

means that a cell has fewer than 30 observations with any gig work and are suppressed; predominantly, these consist

of zeros rather then suppressed data points.

49

Figure A4: County Selection

50

Notes: Counties in dark red had gig platforms enter by 2015. All individuals who become unemployed outside of these counties are excluded from the analysis

sample.

Figure A5: UI Distribution by County

51

Notes: Counts indicate the number of UI events between 2008-2015 by each county.

Figure A6: Illustration of Gig Treatment Intensity Variable

Notes: My ‘Gig Intensity Measure’ captures the number of years gig platforms were available in a given county in

each year and is scaled by 19 —the maximum of years gig platforms were available at the time of job loss across all

individuals in my analysis —to be a measure ∈ [0, 1]. In this example, gig platforms frst enter County A in 2008,

County B in 2010, County C in 2012, and County D in 2014.

52

Figure A7: Linearity of Intensity Measure

Notes: Figure A7 plots the average, across counties, percent of the working age population with any amount of gig

earnings in that county-year by my gig intensity measure. The bottom panel shows the distribution of my gig intensity

measure across individuals in my sample.

53

Figure A8: Predicted Gig Propensities with Respect to Key Predictors

(a) Age

(b) Income

(c) Wages

(d) Individual’s HH Wage Share

Notes: Age is at the time of UI receipt. Income is household AGI from Form 1040 for flers and the sum of information

returns for non-flers (e.g. W-2s, 1099s, etc.). Wage Earnings are the sum of all W-2 wages. Individual household

wage share is the an individual’s wage earnings as a fraction of the sum of the individual’s and spouse’s wages.

54

Figure A9: Distribution of Gig Propensities by Gig Availability

(a) Among High-gig-propensity Individuals

(b) Among Low-gig-propensity Individuals

Notes: High-gig-propensity individuals have predicted propensity values > 0.2 and low-gig-propensity individuals

have predicted propensities between 0.02 and 0.20.

55

Figure A10: Yearly Coeffcients for Household Adjusted Gross Income

(Prime-Age Workers)

Notes: Dependent variable is Household AGI (2017 $), and the top and bottom 1% of values are winsorized. Coeffcient estimates for Titk ∗ Gi ∗ Hi are plotted for each year in event time relative to event time k=-2. These estimates

are interpreted as the effect of having the most gig availability at UI receipt compared to no gig availability among

high-gig-propensity individuals, differencing out any changes that occur among the low-gig-propensity individuals,

and each yearly coeffcient is relative to two years prior to UI receipt. Sample balanced on event time k ∈ [−3, 2].

Regression includes individual FEs, county-by-year FEs, and age FEs. Data drawn around UI recipients frst UI event

in the period 2008-2015 and includes observations from 2005-2017. Robust standard errors clustered by individual.

56

Figure A11: Yearly Coeffcients for Individual Income

(Older Workers)

Notes: Dependent variable is individual income (2017 $), and the top and bottom 1% of values are winsorized.

Restricted to older workers, ages 55-69 at the time of UI receipt. Coeffcient estimates for Titk ∗ Gi ∗ Hi are plotted for

each year in event time relative to event time k=-2. These estimates are interpreted as the effect of having the most gig

availability at UI receipt compared to no gig availability among high-gig-propensity individuals, differencing out any

changes that occur among the low-gig-propensity individuals, and each yearly coeffcient is relative to two years prior

to UI receipt. Sample balanced on event time k ∈ [−3, 2]. Regression includes individual FEs, county-by-year FEs,

and age FEs. Data drawn around UI recipients frst UI event in the period 2008-2015 and includes observations from

2005-2017. Robust standard errors clustered by individual.

57

Figure A12: Robustness to Defnition of High and Low-gig-propensity

(Prime-Age Workers)

(a) Gig Job (x 100)

(b) Individual Income (2017 $)

Notes: In the top panel, the dependent variable is an indicator for having any gig work in a given year, in percentage

points. In the bottom panel, the dependent variable is individual income (2017 $), and the top and bottom 1% of values

are winsorized. Sample is restricted to prime-age workers, ages 25-54. Coeffcient estimates for Titk ∗ Gi ∗ Hi are plotted

for each year in event time relative to event time k=-2. These estimates are interpreted as the effect of having the most

gig availability at UI receipt compared to no gig availability among high-gig-propensity individuals, differencing out

any changes that occur among the low-gig-propensity individuals, and each yearly coeffcient is relative to two years

prior to UI receipt. Sample balanced on event time k ∈ [−3, 2]. Regression includes individual FEs, county-by-year

FEs, and age FEs. Data drawn around UI recipients frst UI event in the period 2008-2015 and includes observations

from 2005-2017. Robust standard errors clustered by individual.

58

Figure A13: Placebo Test for Individual Income (2017 $)

(Prime-Age Workers)

Notes: The dependent variable is individual income (2017 $), and the top and bottom 1% of values are winsorized.

Sample is restricted to prime-age workers, ages 25-54. Coeffcient estimates for Titk ∗ Gi ∗ Hi are plotted for each year in

event time relative to event time k=-2. These estimates are interpreted as the effect of having the most gig availability

at UI receipt compared to no gig availability among high-gig-propensity individuals, differencing out any changes that

occur among the low-gig-propensity individuals, and each yearly coeffcient is relative to two years prior to UI receipt.

Sample balanced on event time k ∈ [−3, 2]. Regression includes individual FEs, county-by-year FEs, and age FEs.

Data drawn around UI recipients frst UI event in the period 2002-2005 and includes observations from 1999-2006.

Robust standard errors clustered by individual.

59

Table A1: Pooled Effects on Labor Supply, Income, and Social Insurance Receipt

(Prime-Age Workers)

VARIABLES

(1)

Gig Year

(x100)

(2)

Gig Earnings

(3)

Individual

Income

(4)

HH AGI

(5)

Wages

(6)

Wage Job

(x100)

Pooled (All Post Years)

-0.0358***

(0.00742)

0.110***

(0.0384)

1.048***

(0.120)

17.27***

(0.752)

-1.408

(0.922)

-49.07***

(4.588)

63.29***

(14.75)

1,948***

(101.4)

-4,728***

(109.6)

64.91

(317.0)

50.52

(410.7)

-853.3

(1,091)

-6,524***

(167.3)

-1,921***

(498.4)

-1,041

(681.3)

-2,025

(1,812)

-13,098***

(132.1)

-565.5

(371.2)

700.8

(507.7)

-2,124*

(1,246)

-3.643***

(0.179)

7.813***

(0.472)

-0.874

(0.700)

-1.921

(1.784)

Observations (Unweighted)

Observations (Weighted)

R-squared

5,586,081

86,843,796

0.246

5,586,081

86,843,796

0.228

5,586,081

86,843,796

0.705

5,004,336

77,627,370

0.763

5,586,081

86,843,796

0.679

5,586,081

86,843,796

0.336

Pre-Period Dep Var Mean

Pre-Period Dep Var SD

0.01

0.77

0.32

93.24

32,036

25,414

45,178

39,679

33,097

29,275

92.8

25.9

Post

Post x Gig Intensity

Post x High

60

Post x Gig Intensity x High

Notes: Results presented are for the subsample of prime-age workers, those ages 25-54 at the time of UI receipt. Post UI, k ≥ 0, indicate years following (and

including) UI receipt. Gig Intensity is a measure, between 0 and 1, of the degree of gig availability in the county in which an individual lives in at the time of

unemployment insurance receipt and is fxed within individual. High is an indicator variable denoting that an individual is in the high predicted gig propensity

sample. Regression includes individual fxed effects, county-by-year FEs, and age FEs. Data drawn around UI recipients frst UI event in the period 2008-2015 and

includes observations from tax years 2005-2017. Top 1% of wages and the top and bottom 1% of income and AGI are winsorized. All dollar values are infation

adjusted to 2017 dollars using CPI-U. Robust standard errors clustered by individual in parentheses (∗∗∗ p < 0.01,∗∗ p < 0.05,∗ p < 0.1).

Table A2: Pooled Effects on Labor Supply, Income, and Social Insurance Receipt

(Older Workers)

VARIABLES

(1)

Gig Year

(x100)

(2)

Gig Earnings

(3)

Individual

Income

(4)

HH AGI

(5)

Wages

(6)

Has SSDI

(x100)

(7)

Has Soc Sec Ret

(x100)

Pooled (All Post Years)

0.0413

(0.0445)

0.341*

(0.189)

0.584

(0.543)

27.44***

(3.564)

13.40*

(7.723)

-46.51

(30.91)

-103.2

(83.12)

4,395***

(563.8)

-4,817***

(537.3)

3,586***

(1,367)

-2,077

(1,689)

5,204

(4,404)

-6,373***

(836.7)

3,457

(2,142)

-2,977

(2,568)

6,000

(7,799)

0.568

-17,384***

(0.411)

(686.5)

-7.749***

637.8

(1.215)

(1,781)

0.418

248.8

(1.247)

(2,336)

-5.896**

8,825

(2.420)

(7,230)

2.020***

(0.556)

1.886

(1.536)

0.664

(1.923)

-2.877

(5.334)

Observations (Unweighted) 303,549

5,000,406

Observations (Weighted)

0.302

R-squared

303,549

5,000,406

0.306

303,549

5,000,406

0.753

273,582

4,504,347

0.803

303,549

5,000,406

0.708

303,549

5,000,406

0.599

303,549

5,000,406

0.791

Pre-Period Dep Var Mean

Pre-Period Dep Var SD

0.48

175.88

42,397

28,916

66,580

48,532

44,222

34,283

1.09

10.38

3.20

17.60

Post

Post x Gig Intensity

= Post x High

61

Post x Gig Intensity x High

0.01

0.82

Notes: Sample is restricted to the older workers subsample, those ages 55-69 at the time of UI receipt. Post UI, k ≥ 0, indicate years following (and including) UI

receipt. Gig Intensity is a measure, between 0 and 1, of the degree of gig availability in the county in which an individual lives in at the time of unemployment

insurance receipt and is fxed within individual. High is an indicator variable denoting that an individual is in the high predicted gig propensity sample. Regression

includes individual fxed effects, county-by-year FEs, and age FEs. Data drawn around UI recipients frst UI event in the period 2008-2015 and includes observations

from tax years 2005-2017. Top 1% of wages and the top and bottom 1% of income and AGI are winsorized. All dollar values are infation adjusted to 2017 dollars

using CPI-U. Robust standard errors clustered by individual in parentheses (∗∗∗ p < 0.01,∗∗ p < 0.05,∗ p < 0.1).

Appendix B

B.1

Data Appendix

The Gig Economy, Form 1099-MISC and Form 1099-K

Prior to the introduction of Form 1099-K in 2011, payments issued by the online gig platforms

would only be found on 1099-MISCs. However, following the introduction of the 1099-K some

platforms chose to start issuing 1099-Ks to report payments to workers.

Complicating matters, there is no consistency across platforms in their decision of which

form(s) to issue to their workers. Some frms may issue only a 1099-MISC while others issue

only a 1099-K, and others may send both. One example of a scenario under which a platform

might issue both forms would be if they reported payments from customers on Form 1099-K and

reported bonuses or other incentive payments on Form 1099-MISC. Thus, I considered payments

that are reported on both 1099-MISC and 1099-K from the list of online platform economy Employer Identifcation Numbers (EINs). The list of online gig platforms consists of approximately

50 labor based platforms and is drawn from online public lists.

1099-MISC: IRS form 1099-MISC reports miscellaneous income. Of particular relevance is income that is reported on line 7, non-employee compensation, which includes payments for services

that an individual provides when the individual is not an employee. This broadly includes all income earned as an independent contractor. However, other sources of income may also be reported

by a payer, which include but are not limited to: rents; prizes and awards; royalties; medical and

health care payments; crop insurance proceeds; fshing boat proceeds. The relevant fling threshold requirement for non-employee compensation is 600$. Thus, for amounts of income earned

above this threshold from a specifc payer, the payer must fll out a 1099-MISC form denoting the

income. The payer submits a copy of the form to both the IRS (directly) and to the payee. The

payer’s name and the payer’s EIN (employer identifcation number) are required on the form, and

thus can be used to distinguish between income earned through platforms classifed as the online

platform economy and other sources.

1099-K: IRS form 1099-K reports Payment Card and Third Party Network Transactions. In 2011,

this form was introduced to increase compliance. Filing is required when total gross payments

exceed $20,000 and 200 transactions. While the threshold for 1099-K is high, many individuals

within our identifed gig population also received 1099-Ks even with income below these thresholds. I present the distribution of 1099-K by dollar amounts among gig workers in Figure B1.

Schedule C: Schedule reports Proft or Loss from Business (Sole Proprietorship). When fling,

individuals self-report their “principal business or profession”. For the small subset of individuals

who do not receive either a 1099-MISC nor a 1099-K, I utilize information in these text strings to

62

identify individuals who work in the online platform economy.

Figure B1: Distribution of 1099 $ Amounts

(a) 1099MISC

(b) 1099K

Notes: This Figure presents the distribution of the dollar amounts that an individual receives on a 1099-MISC (Panel

A) and 1099-K (Panel B) related to gig work. The sample is restricted to individuals who are identifed as Gig

Participants. The width of each bin is $600 (Panel A) and $1000 (Panel B). The fling requirements are indicated with

the red line at 600$ (Panel A) and 20,000$ (Panel B).

63

B.2

Data Cleaning

As the data are arranged and stored with the purpose of tax administration rather than research,

there are several key decisions that have to be made in cleaning the data. First, each individual

or entity is identifed by a Taxpayer Identifcation Numbers (TIN). For individuals this is typically

a social security number (SSN), but may also be an individual taxpayer identifcation number

(ITIN) for non-resident or resident aliens, or even in the case of sole-propietorships an Employer

Identifcation Number (EIN).

EINs typically represent what we think of as a business, but in the case of sole-proprietorships

when there is no legal distinction between the individual and the their business entity. However,

not all sole-proprietors register for an EIN. Thus, they may fle and/or receive their information

returns under either their SSN or their EIN.

B.3

Sample Construction

Among this set of UI recipients, I draw a stratifed random sample based on the last four digits

of an individual’s SSN. I stratify individuals based on whether I ever observe an individual with

any gig work in 2005-2017, and over-sample from the group of ever gig workers since gig work

is an outcome of interest and accounts for a smaller fraction of the individuals in my sample.

Specifcally, I take a 1% random sample of individuals that I never observe taking up gig work

and a 100% sample of individuals that I ever observe with any gig work, regardless of whether it

is in the period around unemployment that I examine. I use sampling weights to account for this

stratifed random sampling methodology in all analyses.

I present weighted population level counts in Table B1 to provide a sense of how each sample

restriction leads to the fnal set of observations. The overall counts restrict to individuals between

the ages of 14-69 at UI receipt, to exclude outlying or potentially erroneous observations and I also

drop all individuals who die during the period three years post-UI, to keep the sample balanced.

From the overall sample of individuals, I split the sample into two sub-groups based on their age

at UI receipt: prime-age workers, ages 25-54, and older workers, ages 55-69. These are the two

key groups that I will focus on in this sample.

Between 2005-2017, there are 68 million UI events experienced by 53 million unique individuals, of which 1,194,819 I observe as ever having any income as a gig worker. The ever gig workers

make up only about 2% of the UI recipients; however, this includes many individuals for whom

gig platforms were not available.

The data do not allow me to differentiate between two separate unemployment shocks and

subsequent UI claims that occur in consecutive years from benefts from one UI claim that span

two calendar years. Thus, I defne an unemployment event as a year in which an individual has

64

positive unemployment compensation in a given year (as reported on Form 1099-G) and zero unemployment compensation in the prior year. I restrict to unemployment events that occur between

2008-2015 in order to have at least three pre- and post-UI event years for each individual. This

drops 14 million individuals from the sample, retaining 39 million unique individuals who experience 49 million UI events. For the 23% of individuals with multiple events, I select the frst event

within this time period.

Finally, I restrict the sample to UI recipients living in counties where gig platforms eventually

enter during the time period of analyzed UI events (2008-2015). This excludes counties where gig

platforms never enter or where the frst gig platforms had not yet entered as of 2015.43 Appendix

Figure A4 identifes the 819 selected counties out of 3,021 US counties. As seen in Appendix

Figure A5, the selected counties contain the majority of UI claims. This sample restriction retains

about 83% of all UI recipients.

43 I infer the availability of gig platforms based on individual level data stemming from the universe of individuals in a

county. See Appendix Section 2.2 for more information.

65

Table B1: Sample Restrictions: Numbers of Observations

66

All Ages

Sample

Prime-Age

Near-Elderly

Gig Workers

(All Ages)

All UI Events 2005-2017

Number of UI Events

Unique Individuals

...with 1 UI event

...with 2+ UI events

68,510,061

53,073,619

40,087,646

12,985,973

91,340,712

38,827,643

27,446,153

11,381,490

17,877,935

8,277,077

7,087,937

1,189,140

1,623,961

1,194,819

839,646

355,173

Restricting to First UI Event 2008-2015

Number of UI Events

Unique Individuals

48,644,065

38,714,964

36,713,725

28,003,976

7,052,173

6,161,833

1,176,865

886,164

Unique Individuals (sample restrictions)

First UI Event 2008-2015

Drop non-US counties

Drop never Gig Areas

Drop gig after 2015 areas

38,714,964

38,714,964

34,956,717

32,217,165

28,003,976

27,965,010

25,343,612

23,375,773

6,161,833

6,157,708

5,565,168

5,131,582

886,164

885,451

875,217

863,565

Final Stratifed random sample

Unique Individuals

1,177,101

917,128

106,243

863,565

Notes: This table provides population weighted counts for the overall sample and the two sub-groups focused on in this paper. Prime-age workers are ages 25-54

at UI receipt. Older workers are those ages 55-69 at the time of UI receipt. Each row denotes the number of observations or unique individuals in each sample

restriction.

Appendix C

Take-up of Unemployment Insurance

The purpose of this section is to test if the rollout of gig platforms affected the take-up rate of UI

benefts. Using publicly available data from the Department of Labor Employment and Training

Administration on the number of unemployed and insured unemployed at the state-quarter level, I

estimate the following regression equation:

InsuredUnemployedst =α + β1 TotalUnemployedst ∗ GigUnavailablest

+ β2 TotalUnemployedst ∗ GigAvailablest + ηs + γt + εst

(4)

I aggregate the county-level availability, described in Section 2.2, to the state-level by taking

the earliest year of gig availability year across all counties in a state. Since this is a much noisier

approximation of gig availability across counties within a state, I simply present the coeffcient as

a pre- versus post-gig availability rather than approximating gig intensity based on the number of

years. Figure C1 presents the regression coeffcients visually. Accounting for year and state fxed

effects, there was a take up rate of UI in states and years where gig platforms were available and

were not of 28% and 28.6%, respectively. Estimates of β1 and β2 are not statistically different from

one another —I present standard errors in parantheses.

While gig availability does not appear to affect individuals decision to take up UI benefts, it

may affect the duration of UI benefts and/or the amount of benefts that an individual receives. I

examine this in Figure C2. While I do not fnd a statistically signifcant decrease in the total amount

of annual unemployment compensation that individuals receive, there is suggestive evidence that

those with gig availability have slightly lower annual receipt. This suggests that those individuals

with gig availability are either staying on UI for a shorter duration or may be receiving lower

benefts as earnings from temporary work reduces the amount of benefts for which an individual

is eligible to receive.44

Tables C1 and C2 present data on the characteristics of unemployment insurance applicants and

recipients from the Bureau of Labor Statistics (BLS). Table C1 provides summary characteristics

on individuals by whether they apply for UI insurance to highlight differences and similarities

between these two groups. Those who applied for UI in 2018 tended to be older, were more likely

to be male, and were more likely to hold a professional certifcation or license. Table C2 provides

summary statistics on potential reasons why individuals do no apply for UI benefts. The most

common reason given, by 61% of respondents, was eligibility issues. The second most common

44 The exact beneft formulas and how much income an individual can earn depends on each state’s rules.

Though in

many states individuals may work part-time and receive reduced benefts. Therefore there is a certain threshold at

which their benefts are completely offset based on their weekly earnings. Income received from contract and selfemployment work is considered in addition to wages in determining how much to reduce benefts.

67

reason was the respondents expected to start working again soon.

Figure C1: Take-Up Rates of Unemployment Insurance (UI)

Notes: Data from Department of Labor Employment and Training Administration. Data are at the quarterly by state

level, and include the total number of unemployed individuals and the number of insured unemployed. Plot controls

for state and year FEs.

68

Figure C2: Effects on UI Compensation Amount - Duration Effects

Notes: Dependent variable is the amount of unemployment compensation in (2017 $) an individual received (Form

1099-G), and the top 1% of values are winsorized. Coeffcient estimates for Titk ∗ Gi ∗ Hi are plotted for each year in

event time relative to event time k=-2. These estimates are interpreted as the effect of having the most gig availability

at UI receipt compared to no gig availability among high-gig-propensity individuals, differencing out any changes that

occur among the low-gig-propensity individuals, and each yearly coeffcient is relative to two years prior to UI receipt.

Sample balanced on event time k ∈ [−3, 2]. Regression includes individual FEs, county-by-year FEs, and age FEs.

Data drawn around UI recipients frst UI event in the period 2008-2015 and includes observations from 2005-2017.

Robust standard errors clustered by individual.

69

Table C1: Characteristics of UI Applicants vs Non-Applicants

70

Table C2: Reasons for Not Applying for UI Benefts

Source:

Characteristics of Unemployment

https://bls.gov/news.release/pdf/uisup.pdf

Insurance

71

Applicants

and

Beneft

Recipients

¯

2018.

Appendix D

Changes in Education Decisions

In this appendix, I examine how the availability of gig platforms affects individuals decision to

attend a post-secondary institution. For example, workers may forego additional education or

vocational training following an unemployment shock in exchange for earning income through the

gig economy. Alternatively, the fexibility that gig work provides may allow more individuals to

go back to school following job loss when they might not otherwise have been able to. I estimate

Equation 3 with an indicator variable for being a post-secondary student in a given year as the

outcome variable, and present the results in Appendix Figure D1. At least for prime-age workers,

I fnd no evidence of changes in education decisions among high-gig-propensity individuals with

more gig availability.

Figure D1: Yearly Coeffcients for Being a Student

Notes: Dependent variable is an indicator for being a student in a given year as identifed by having an eligible tuition

payment on Form 1098-T (in percentage points 0 or 100). Coeffcient estimates for Titk ∗Gi ∗Hi are plotted for each year

in event time relative to event time k = −2. Sample balanced on event time k ∈ [−3, 2]. Regression includes individual

FEs, county-by-year FEs, and age FEs. Data drawn around UI recipients frst UI event in the period 2008-2015 and

includes observations from the tax years 2005-2017. Robust standard errors clustered by individual.

72

Appendix E

Social Security Withdrawals

In this appendix, I zoom in to those ages 62-67 at the time they face their unemployment shock as

these are the subset of individuals who are able to respond on this margin. I estimate a comparable

set of specifcations to those found in Figure 9 and in column 7 of Tables 7, 8, and A2.

Figure E1: Yearly Coeffcients for Social Security Retirement Withdrawals (Ages 62-67)

Notes: Dependent variable is an indicator having received social security disability income benefts (in percentage

points 0 or 100). Restricted to a subset of older workers, those ages 62-67 at UI receipt. Coeffcient estimates for Titk ∗

Gi ∗Hi are plotted for each year in event time relative to event time k=-2. These estimates are interpreted as the effect of

having the most gig availability at UI receipt compared to no gig availability among high-gig-propensity individuals,

differencing out any changes that occur among the low-gig-propensity individuals, and each yearly coeffcient is

relative to two years prior to UI receipt. Sample balanced on event time k ∈ [−3, 2]. Regression includes individual

FEs, county-by-year FEs, and age FEs. Data drawn around UI recipients frst UI event in the period 2008-2015 and

includes observations from 2005-2017. Robust standard errors clustered by individual.

73

Table E1: Results—Older Workers - Social Security Withdrawals

(1)

(2)

(3)

Has Soc Sec Ret (x 100)

VARIABLES

Post

9.916***

(2.879)

8.345

(8.569)

21.23**

(9.419)

-28.68

(18.94)

Post x Gig Intensity

Post x High

Post x Gig Intensity x High

26.91***

(3.701)

8.774

(7.577)

5.393

(5.325)

-12.06

(13.97)

11.65***

(2.767)

3.073

(6.168)

7.634

(6.142)

-13.84

(15.02)

Short Run (First Post Year)

Long Run (Two-Four Years Post)

Pooled (All Post Years)

Observations (Unweighted)

Observations (Weighted)

R-squared

X

31,723

531,376

0.791

42,789

734,799

0.877

X

57,196

973,144

0.807

Pre-Period Dep Var Mean

Pre-Period Dep Var SD

7.25

25.93

7.25

25.93

7.25

25.93

X

Notes: Results presented for a subset of older workers, those ages 62-67 at UI receipt. Post UI, k ≥ 0, indicate years

following (and including) UI receipt. Gig Intensity is a measure, between 0 and 1, of the degree of gig availability in

the county in which an individual lives in at the time of unemployment insurance receipt and is fxed within individual.

High is an indicator variable denoting that an individual is in the high predicted gig propensity sample. Regression

includes individual fxed effects, county-by-year FEs, and age FEs. Data drawn around UI recipients frst UI event

in the period 2008-2015 and includes observations from tax years 2005-2017. Robust standard errors clustered by

individual in parentheses (∗∗∗ p < 0.01,∗∗ p < 0.05,∗ p < 0.1).

74