Full text of Working Papers (Federal Reserve Bank of Chicago) : Employee Costs of Corporate Bankruptcy, Working Paper 2022-09

The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.

Federal Reserve Bank of Chicago

Employee Costs of Corporate Bankruptcy
John R. Graham, Hyunseob Kim, Si Li, and
Jiaping Qiu

March 2022
WP 2022-09
https://doi.org/10.21033/wp-2022-09
*

Working papers are not edited, and all opinions and errors are the
responsibility of the author(s). The views expressed do not necessarily
reflect the views of the Federal Reserve Bank of Chicago or the Federal
Reserve System.

Employee Costs of Corporate Bankruptcy*
John R. Graham, Hyunseob Kim, Si Li, Jiaping Qiu †
March 2022
Abstract
An employee’s annual earnings fall by 13% the year her firm files for bankruptcy, and the present
value of lost earnings from bankruptcy to six years following bankruptcy is 87% of pre-bankruptcy
annual earnings. More worker earnings are lost in thin labor markets and among small firms. Ex
ante compensating wage differentials for this “bankruptcy risk” are approximately 2% of firm
value for a firm whose credit rating falls from AA to BBB, comparable to the magnitude of debt
tax benefits. Thus, wage premia for expected costs of bankruptcy are of sufficient magnitude to be
an important consideration in corporate capital structure decisions.
JEL Codes: G32, G33, J21, J31, J61.
*

We thank the Editors (Stefan Nagel and Ken Singleton), an anonymous associate editor, and two anonymous referees for
valuable comments and support, and John Abowd, Ashwini Agrawal, Jonathan Berk, Rob Fairlie, Todd Gormley, Michael
Hertzel, Joseph Hotz, Po-Hsuan Hsu, Robert Jarrow, Yawen Jiao, Pab Jotikasthira, Andrew Karolyi, Bryant Kim, Pauline
Leung, Chen Lin, David Matsa, Roni Michaely, William Mullins, Justin Murfin, Paige Ouimet, Wendy Rotenberg, Daniel
Sullivan, Yongjun Tang, Karin Thorburn, Michael Waldman, Wei Wang, Liu Yang, Hayong Yun, and seminar and
conference participants at AFA, Census Bureau RDC Conference, Chicago Fed, Chinese University of Hong Kong, CICF,
City University of Hong Kong, Cornell University (Finance and Labor Economics), CSEF-EIEF-SITE Conference on
Finance and Labor, EFA, FIRS Conference, FMA Asia, IDC, Maryland Junior Corporate Finance Conference, NFA, SFS
Finance Cavalcade, SOLE/EALE, University of Calgary, University of Hong Kong, and U.S. Census Bureau (CES) for
helpful feedback. We also thank Bert Grider at the TFSRDC for help with data and clearance requests, Nichole Szembrot at
the NYFSRDC (Cornell) for administrative support, and Tanya Bakshi, Vitor Costa, Daniel Gallego, Dawoon Kim, Eric
Kim, Cindy Lu, and Katya Potemkina for research assistance. Any views expressed are those of the authors and not those
of the U.S. Census Bureau. The Census Bureau’s Disclosure Review Board and Disclosure Avoidance Officers have
reviewed this information product for unauthorized disclosure of confidential information and have approved the disclosure
avoidance practices applied to this release. This research was performed at a Federal Statistical Research Data Center under
FSRDC Project Number 808 (CBDRB-FY21-P808-R8890 , FY22-P808-R9448). This research uses data from the Census
Bureau’s Longitudinal Employer Household Dynamics Program, which was partially supported by the following National
Science Foundation Grants SES-9978093, SES-0339191 and ITR-0427889; National Institute on Aging Grant AG018854;
and grants from the Alfred P. Sloan Foundation. We are grateful to Edward Altman at NYU for sharing his dataset of default
events, Antonio Falato for sharing DealScan-Compustat merged data and related programs, Lynn LoPucki at UCLA for
sharing his Bankruptcy Research Database, and David Matsa for sharing data on state-level UI generosity. Kim
acknowledges generous financial support from the Kwanjeong Educational Foundation. Li and Qiu acknowledge financial
support from the Social Sciences and Humanities Research Council of Canada. This paper was previously circulated under
the title “Human Capital Loss in Corporate Bankruptcy.”
†
John R. Graham (john.graham@duke.edu), Fuqua School of Business, Duke University, Durham, NC 27708, phone: (919)
660-7857; Hyunseob Kim (hyunseob.kim@chi.frb.org), Economic Research Department, Federal Reserve Bank of Chicago,
Chicago, IL 60604, phone: (312) 322-6131; Si Li (sli@wlu.ca), Financial Services Research Centre, Lazaridis School of
Business and Economics, Wilfrid Laurier University, Waterloo, Ontario N2L 3C5, Canada, phone: (519) 884-0710 ext. 2395;
Jiaping Qiu (qiu@mcmaster.ca), DeGroote School of Business, McMaster University, Hamilton, Ontario L8S 4M4, Canada,
phone: (905) 525-9140 ext. 23963.

1. Introduction
When a firm experiences a substantial negative shock, its employees bear personal costs
resulting from labor market adjustments. An extensive literature documents such costs to workers
caused by plant closures, international trade, and environmental regulation, among other shocks
(Jacobson, LaLonde, and Sullivan, 1993; Walker, 2013; Autor et al., 2014; Hummels et al., 2014).
In this paper, we focus on costs to the workforce resulting from corporate bankruptcy and the
implications for the firm’s risk-sharing role and capital structure decisions.
According to the theory of the firm as an insurance provider, it is optimal for a risk-neutral
firm to provide insurance to risk-averse workers (e.g., Baily, 1974; Azariadis, 1975). Workers are
willing to accept lower wages, implicitly paying for this insurance (Guiso, Pistaferri, and Schivardi,
2005). When firms choose to use more debt financing (perhaps to achieve tax benefits), they
increase bankruptcy risk, thereby reducing the risk-sharing benefit provided to employees (Berk,
Stanton, and Zechner, 2010). Employees in response demand higher promised wages (Abowd and
Ashenfelter, 1981). We estimate these compensating wage differentials, which can be thought of
as an implicit bankruptcy-related cost of debt (Agrawal and Matsa, 2013). Most empirical
estimates of the bankruptcy costs of debt seem small relative to the benefits (the “debt
conservatism” puzzle; Graham, 2000). 1 In this vein, our estimates help quantify a hard-to-measure
indirect cost of bankruptcy, namely increased wages due to reduced risk sharing. Our analysis can
thus be viewed as a step towards the goal of understanding the role of financial distress-induced
employee costs in capital structure decisions.
To estimate the impact of an employer’s bankruptcy filing on its workers’ earnings, we
rely on worker-firm matched data from the U.S. Census Bureau’s Longitudinal EmployerHousehold Dynamics (LEHD) program, combined with a comprehensive database of public firm
bankruptcy filings. The LEHD data allow us to follow individual workers across firms over time
and observe their earnings and other characteristics of employment, such as industry and

1

Traditional costs of financial distress or bankruptcy include those due to the bankruptcy process, such as lawyer’s
fees, loss of market share, and asset fire sales.

1

geographical location. As a result, we are able to estimate the effect of corporate bankruptcy on
employees, even when they change employers post-bankruptcy.
We construct a sample of 130 bankruptcy filings by U.S. public firms from 1992 to 2005
and follow for up to six years approximately 234,000 workers who were employed by the bankrupt
firms one year before bankruptcy. We estimate the effect of a bankruptcy filing on worker
outcomes by comparing affected employees to workers in control firms matched using key firm
characteristics before bankruptcy (except for financial leverage), such as the firm’s industry,
economic performance, and changes in the market value of assets. In addition, we control for
unobserved heterogeneity in earnings across industries and geographies over time, and workers,
using detailed fixed effects such as industry × county × year and worker fixed effects. Hence, our
empirical approach compares within-worker earnings dynamics of employees of bankrupt and
control firms with similar observables in the same industry, county, and year. Although it may not
be possible to completely control for all unobservables that affect selection into bankruptcy filing,
the comparison conditional on matched firm characteristics and detailed fixed effects helps
mitigate many potentially confounding factors.
We find that on average employee earnings begin to deteriorate in the year of bankruptcy,
relative to those of the control group. One year after bankruptcy, annual earnings are 13%
(adjusting for unemployment insurance (UI)) to 14% (without UI adjustment) lower than prebankruptcy annual earnings (see Figure 1). The present value (PV) of cumulative earnings losses
from the year of bankruptcy to six years afterward is 87% (with UI adjustment) to 89% (without
UI adjustment) of pre-bankruptcy annual earnings. Importantly, earnings losses are larger for
workers who leave the industry post-bankruptcy (suggestive that loss of human capital plays a role
in worker earnings loss following corporate bankruptcy), and who worked in thinner local labor
markets before bankruptcy (suggestive that labor mobility in thick labor markets helps attenuate
earnings losses following corporate bankruptcy).
In competitive labor markets, employees would consider expected losses in their personal
earnings that follow corporate bankruptcy and demand a wage differential ex ante. From the firm’s
perspective, such a wage premium is a cost associated with financial distress risk. Estimating the

2

magnitude of wage premium thus may shed light on the debt conservatism puzzle. We estimate
this compensating wage differential using two approaches that, despite different assumptions and
estimates, produce quantitatively similar magnitudes of wage premia for bankruptcy risk. First,
using the information on employees who move between firms, we regress movers’ highest
quarterly earnings in the last four quarters right before and in the first four quarters right after the
movement on the leverages of their old and new firms, respectively. We find that workers’ wages
increase when they move from firms with low leverage to those with high leverage, controlling for
unobserved and observed firm and worker characteristics, and time-varying industry and local
labor market conditions. The regression coefficient on market leverage allows us to estimate the
wage premium due to an increase in financial distress risk. One caveat of this approach is that
these wage premium estimates may be biased due to omitted variables.
Therefore, we use a second, alternative approach based on a valuation model with plausible
assumptions (e.g., risk-neutral valuation given risk averse workers); we find that the expected
additional present value cost to compensate for bankruptcy-driven earnings loss risk is about 1.0%
(3.0%) of firm value for a typical AA-rated (BBB-rated) firm, relative to a risk-free firm. These
indirect distress costs are substantial in that they amount to 42% (59%) of the tax benefits of
corporate debt for an AA-rated (BBB-rated) firm estimated in previous research (e.g., Graham,
2000). Overall, our estimated employee-related bankruptcy costs are of sufficient magnitude to
potentially help explain the debt conservatism puzzle.
Our paper is related to two strands of literature. The first literature examines the earnings
losses of displaced workers. A central finding of this line of research is that job displacement (i.e.,
job loss resulting from mass layoffs) leads to large and persistent earnings losses (e.g., Jacobson,
Lalonde, and Sullivan, 1993; von Wachter, Handwerker, and Hildreth, 2009; Couch and Placzek,
2010; von Wachter, Song, and Manchester, 2011). This literature also shows that the effect of job
displacement on earnings is especially large when labor market conditions are weaker, such as in
recessions (Schoeni and Dardia, 2003; Kodrzycki, 2007; Davis and von Wachter, 2011).
Our paper differs from the existing work on displaced workers in several important ways.
First, we focus on one specific financial shock on workers, namely corporate bankruptcy, as

3

opposed to pure job displacements. As we show later, the workers in our sample of bankrupt firms
are not necessarily displaced, with a substantial fraction of them staying with the bankrupt
employer. Thus, our estimates of earnings losses are based on bankrupt firm employees, whether
they separate from or remain with the firm after bankruptcy. Second, crucially, it is not possible to
directly infer the employee cost of bankruptcy from estimates in the displacement literature, which
often uses employees who stay with the firm (i.e., firm stayers) as the counterfactual group.
Moreover, relative to workers who leave the firm, firm stayers may have different earnings loss
patterns associated with bankruptcy, which a displaced worker analysis would not capture. Third,
our estimates account for state-level unemployment insurance benefits, which offset part of
earnings losses due to bankruptcy. Fourth, in a step beyond most of the job displacement literature,
we attempt to identify the effect of financial distress on employee earnings using a regression
discontinuity design for loan covenant violations.
The second strand related to our paper is the finance literature on the impact of financial
distress on employees. Our paper is among the first to use worker-firm matched micro data to
examine labor market outcomes for individuals after financial distress of their employer. 2 While
previous research has examined the effects of bankruptcy filings on firm-level outcomes such as
accounting performance and management turnover (Gilson, 1989; LoPucki and Whitford, 1992),
relatively little is known about the consequences of bankruptcy and financial distress for
employees, other than firm-level employment. 3 Our paper estimates detailed changes in employee
earnings following bankruptcy relative to a control group using worker-level data. Importantly,
our estimates have important implications for capital structure theory (see, e.g., Matsa (2018)).
We derive wage premia for bankruptcy risk as labor-related indirect bankruptcy costs.
Capital structure theories model that employee wage loss due to financial distress discourages
2

A contemporaneous paper by Baghai, Silva, Thell, and Vig (2021) uses Swedish worker-firm matched data to
document that financially distressed firms lose their most skilled employees. However, they do not examine changes
in employee earnings around financial distress nor wage premia for financial distress risk, which are our focus.
3
For instance, using Compustat data on firm-level employment, Falato and Liang (2016) show that firms cut
employment substantially after loan covenant violations, and Agrawal and Matsa (2013) show that firms reduce
employment by about 30% during the years surrounding bond defaults. Similarly, Hotchkiss (1995) shows that firms
suffer employment reductions around Chapter 11 bankruptcy filings.

4

firms from using debt (e.g., Titman, 1984; Butt-Jaggia and Thakor, 1994; Berk, Stanton, and
Zechner, 2010). Agrawal and Matsa (2013) and Kim (2020) provide evidence of the importance
of this mechanism by showing that changes in the cost of job loss affect corporate capital structure
decisions. In our paper, we not only provide evidence for the importance of this mechanism, but
also quantify the magnitude of these wage premia for employees’ earnings loss risk due to
bankruptcy. Differing from existing estimates of wage premia based on firm-level wage data
(Chemmanur, Cheng, and Zhang, 2013) or implied by compensating differentials for
unemployment risk (Agrawal and Matsa, 2013), we use worker-level wage data from LEHD to
directly estimate the wage premia. Our analysis suggests that the resulting additional compensation
cost is comparable in magnitude to the tax benefit of debt and thus could play a first-order role in
corporate capital structure decisions.

2. Data and Descriptive Statistics
2.1. Data Sources and Sample Selection
2.1.1. Bankruptcy Event Data and LEHD Datasets
We begin by identifying corporate bankruptcy cases from the UCLA-LoPucki Bankruptcy
Research Database (BRD). 4 The BRD contains public companies with more than $100 million of
assets (measured in 1980 dollars) that filed bankruptcy cases from October 1, 1979 to present. 5
We focus on public firms’ bankruptcy filings because we examine capital structure implications
later in the paper.6

4

We thank Lynn M. LoPucki at UCLA for sharing this database, which is also used by Jiang, Li, and Wang (2012),
Eckbo, Thorburn, and Wang (2016), and Goyal and Wang (2017).
5
The majority of bankruptcy cases in the database are filed under Chapter 11 (reorganization) of the U.S. Bankruptcy
Code, while only a handful are filed under Chapter 7 (liquidation).
6
Using a 235 private firm sample selected by BankruptcyData.com that have public debt or that BankruptcyData
deemed important, we find that employees of these private bankrupt firms experience about three times larger earnings
losses after a bankruptcy filing than their public firm counterparts (see Online Appendix Table A5, column (1)).
Possible explanations include that public firms, which tend to be larger, offer better opportunities for employees to
develop general skills or to switch jobs. The results are consistent with our findings that there is significant
heterogeneity in employee costs of bankruptcy depending on firm size and search frictions in the labor market. See
Bernstein, Colonnelli, and Iverson (2019) for a study of asset reallocation effects of bankruptcy using a large sample

5

We merge these bankruptcy events to worker-firm matched information from the U.S.
Census Bureau’s Longitudinal Employer-Household Dynamics (LEHD) program, establishmentlevel databases, and the Compustat database. The LEHD data we have access to cover 30
participating U.S. states largely from 1990 (coverage in Maryland begins in 1985) to 2008 (see
Table 1 for the list of covered states and years available) and provide detailed information on
employment relationships such as employee earnings, industries, and geographical locations, and
worker characteristics such as age, imputed education, and gender. 7 The annual earnings records
in LEHD are non-missing as long as an individual reports positive earnings in any of the four
quarters in a given year. However, if the individual earns nothing in all four quarters of the year,
the annual record is missing (as opposed to showing zero earnings). Thus, missing annual
observations are due to two potential reasons. First, an individual may move to any of the 20 nonLEHD states or become self-employed (in which case no earnings are reported in the LEHD data).
Second, the individual may become unemployed and earn no wages in a given year. We
empirically address how these sample attrition issues might affect our estimates in Section 3.1.2.
[Table 1 about here]
We link datasets from the LEHD infrastructure with other Census Bureau establishmentlevel datasets and subsequently with Compustat and BRD. Specifically, among the databases
available from the LEHD infrastructure, we use the Individual Characteristics Files (ICF), which
provides worker-level characteristic variables; the Employment History Files (EHF), which
contains annual and quarterly earnings records, locations (state and county) and industries for each
worker-firm pair; and the Unit-to-Worker Imputation File (U2W), which is used for job-location
imputation at the SIC (or NAICS) industry and county level. We use the Business Register Bridge
(BRB) and the Compustat-SSEL Bridge (CSB) in conjunction with the Standard Statistical

of bankruptcy filings by both public and private firms. They focus on the difference between reorganization and
liquidation but do not study employee costs or capital structure implications of the bankruptcy processes.
7
The Census Bureau imputes the education variable as follows. For each state, a logit model is used to estimate the
probability of belonging to one of 13 education categories (e.g., high school, college) using the characteristics such as
age categories, earnings categories, and industry indicators. The education category is then imputed based on the
predicted probability.

6

Establishment List (SSEL) to link the LEHD files with the Census Bureau’s Longitudinal Business
Database (LBD), Compustat and the BRD. To avoid complications associated with legal ages for
employment and early retirement, we require that worker ages be between 22 and 55 in the year
before a bankruptcy filing. To exclude earnings changes due to unstable employment relations
with the firm (e.g., temporary workers), we focus on workers with at least two years of tenure with
the bankrupt firm one year before its bankruptcy filing.8
Because earnings are a key variable in our analysis, we provide details on LEHD earnings
records below. As discussed in Abowd et al. (2009), the LEHD earnings are extracted from the
state unemployment insurance (UI) records and correspond to the report of an individual’s UIcovered earnings, which is retained in the database as long as the worker earns at least one dollar
of UI-covered earnings during a given quarter. According to the Bureau of Labor Statistics (BLS),
UI coverage is comprehensive and comparable across states. For example, UI covered 96% of total
jobs and 92.5% of the wage component of national income in 1994. The UI earnings include gross
wages and salaries, bonuses, stock options, tips and other gratuities, and the value of meals and
lodging, where supplied. In some states, employer contributions to certain deferred compensation
plans, such as 401(k)s, are included in total earnings. 9
2.1.2. Census Establishment-Level Data
We use the Longitudinal Business Database (LBD) to obtain comprehensive data on
employment and payroll at business establishments. The LBD tracks more than five million
establishments each year from 1977 to 2011 covering all U.S. states. The variables available in the
database include employment, annual payroll, industry classifications, location identifiers (e.g.,
states, counties), and parent firm identifiers. Given the comprehensive nature, we use the LBD to
measure firm-level employment, payroll, and the size of local labor markets. However, the LBD
does not contain data on worker-level earnings, industries, or location of employment, which are
contained in the LEHD.
2.1.3. Sample Selection: Bankrupt and Matched Firms
8
9

Requiring at least six years of tenure gives a qualitatively similar result (see Online Appendix Table A5, column (2)).
See https://www.bls.gov/opub/hom/cew/pdf/cew.pdf at the BLS.

7

We begin with 510 bankruptcy cases from the BRD from 1990 to 2005 (excluding
financials and utilities). We match these firms with the SSEL, LBD, and Compustat and require
that they have non-missing employment information from the LBD and financial information from
Compustat. This step produces 360 bankruptcy cases. We then match these bankrupt firms with
LEHD and keep firms with at least ten employees in LEHD one year before filing. Because the
LEHD earnings data we have access to are only available up to 2008, we include only the bankrupt
firms that file for bankruptcy in or before 2005. This restriction allows employees of bankrupt
firms in 2005 to have up to three years of earnings data post-bankruptcy. After matching with
LEHD, we have 190 bankruptcy cases.
To construct a matched sample from non-bankrupt firms, we first require that the nonbankrupt firms have non-missing matching information and have at least ten workers from LEHD
one year before matching from 1990 to 2005. Then for each year and three-digit SIC industry, we
use a propensity-score matching approach in which we use market value of assets (MVA), the log
change in MVA over the previous year, return on assets (ROA), market-to-book, and total number
of employees (from the LBD) to match bankrupt to non-bankrupt firms. More specifically, we
obtain the financial information in the latest fiscal year within two years prior to bankruptcy filings
for bankrupt firms and the information for potential matches in the same industry and year. 10 The
propensity score matching is performed within each industry-year cell to control for the influences
of unobserved time-varying aggregate and industry conditions. Note that we use ROA, market-tobook, and (changes in) MVA but not leverage in matching, which enables us to control for the
impact of (forward-looking) firm performance but also allows bankrupt firms to have different
debt usage than matched firms.
For each bankrupt firm in the sample, we choose one non-bankrupt firm that has the closest
linearized propensity score (LPS or log odds ratio) as a matched firm. We require that the absolute
difference in the LPS be less than or equal to 0.25 (see e.g., Imbens and Rubin, 2015). If we do
not find a matched firm within a given three-digit SIC-year cell based on the cut-off, we repeat the
10

When we conduct the matching based on data from four years prior to bankruptcy filings (which is our benchmark
year in earnings regressions), we obtain similar results (see Online Appendix Table A1).

8

matching process to find a matched firm within a given two-digit SIC-year cell, and then within a
one-digit SIC-year cell. This one-to-one matching procedure gives us a final event sample
consisting of 130 bankrupt and 130 matched firms from 1992 to 2005 that employ at least ten
workers in LEHD one year before bankruptcy.
2.2. Descriptive Statistics
2.2.1. Firm Characteristics of Bankrupt and Control Firms
Table 2 presents descriptive statistics on bankruptcy events, firms and matched firms. Panel
A shows that the distribution of bankrupt firms with employees in the LEHD data is similar to that
of the full sample of bankrupt firms from the BRD, in terms of bankruptcy outcomes. Bankruptcy
events that lead to merger, acquisition, or continuation of the firm represent about 30-40% of the
events in both samples, and those leading to liquidation, firm closure, or refiling represent about
20%. The first four columns of Panel B further investigate the representativeness of the bankrupt
firms that have workers in LEHD based on firm characteristics measured in the latest fiscal year
within two years before bankruptcy. The means show that bankrupt firms from BRD (that have
information from Compustat and LBD; column (1)) and the bankrupt firms with employees in
LEHD (column (3)) are similar in most firm characteristics including firm size (measured using
sales, assets, and number of employees), book and market leverage, total wage bills scaled by
assets, and R&D expense scaled by assets (see column (7) for t-statistics for the differences). The
exceptions are that bankrupt firms in LEHD are significantly more profitable, have lower marketto-book, and have lower capital expenditures relative to net property, plant, and equipment.
Importantly, to verify that the bankrupt firms in our final sample are similar to the matched
firms, Panel B compares the mean and standard deviation characteristics of bankrupt firms with
employees in LEHD (columns (3) and (4)) to propensity-score matched firms (columns (5) and
(6)).11 The t-statistics reported in column (8) on the differences between the means show that the
bankrupt LEHD firms and matched firms do not differ significantly in any of the matching
variables, including the level and log change in market value of assets, the number of employees,

11

The Census Bureau does not permit researchers to disclose percentiles such as the median.

9

ROA, and market-to-book. The two groups also have similar sales, book assets, wages-to-assets
ratios, capital expenditures, and R&D intensity, despite these variables not being explicitly
employed in matching. Notably, the bankrupt firms have significantly higher book and market
leverage than matched firms (60% vs. 35% for book leverage and 57% vs. 32% for market
leverage), which is consistent with leverage playing a role in “treated” firms ending up in
bankruptcy. (Recall that the matching process intentionally does not match on leverage.)
[Table 2 about here]
To check the similarity of the pre-event trends between the bankrupt and matched samples,
Table 2, Panel C examines the dynamics of firm characteristics for the two samples starting three
years before bankruptcy filings (see Online Appendix Figure A1 for a visual presentation of the
results). To facilitate a balanced comparison of statistics across the years, we focus on a subset of
bankrupt firms that have complete financial information from Compustat and LBD (on
employment and payroll) during all the three years t-3, t-2, and t-1. This requirement leaves us
with 90 bankrupt firms and their corresponding control firms. 12 The panel shows that each year
during the pre-event window, all variables but leverage and log market value of assets in t-2 have
similar values for treated and matched firms, indicating parallel trends. Importantly, though the
(log) market value of assets for bankrupt firms declines over the years, the difference in cumulative
changes in log from t-3 between the treatment and control groups is not statistically significant.
Notably, both groups of firms experience an upward trend in leverage prior to bankruptcy, with
bankrupt firms exhibiting substantially higher leverage than matched firms in each year before
bankruptcy.
Overall, the results in Table 2 confirm that bankrupt and matched firms are similar in
important firm characteristics including market value, growth opportunities, profitability, and their
trends prior to the event year, except for leverage, suggesting that the matched firms are a valid
control sample for our analysis.13

12

We find similar magnitudes of PV losses for this subset of events (see Online Appendix Table A5, column (4)).
Of course, we cannot perfectly ensure that leverage is the only difference between the bankrupt and control firms.
Thus, we interpret our results with caution (see also Section 3.1).
13

10

2.2.2. Employee Characteristics of Bankrupt and Control Firms
Table 3 presents employee characteristics from LEHD for the bankrupt and control firms.
We examine employees of these firms one year prior to the bankruptcy filing. The worker
characteristics, including imputed education, age, experience, gender, and earnings, are similar
between the employees of bankrupt and the matched firms. All of the differences are insignificant
at a conventional level (column (7)). In addition to comparing with the propensity-score matched
sample, we follow the literature on labor market adjustments to job displacements (e.g., Couch
and Placzek, 2010; Davis and von Wachter, 2011) and construct an alternative, less strict control
sample of employees. To circumvent computational constraints, we randomly select 0.5% of
workers in the LEHD universe (which contains more than 2.8 billion quarterly earnings records)
who are employed by non-bankrupt public firms. We impose the same requirements on the control
group for industry (i.e., exclude financial and utilities sectors), tenure, and age as we impose on
the sample of bankrupt firms (see Section 2.1.1). Column (8) shows that all characteristics are
similar between the bankruptcy and the randomly selected workers. Overall, the comparisons of
employee characteristics suggest that both the propensity-score matched and randomly selected
groups of workers are comparable to the treatment group in terms of observable characteristics.
The bottom half of Table 3 documents the rate at which employees move across firms,
industries, and counties around bankruptcy filings. We first report the percentage of workers who
are employed by the firm in year t-4 and stay in the same firm, industry, or county through t-1,
separately for treated and matched firms. The results show that before bankruptcy, the treated and
matched control groups are comparable in terms of employee staying. In particular, the rates of
staying in the same firm, industry, and county and remaining in LEHD from t-4 to t-1 are 54.6%
vs. 51.4%, 63.7% vs. 63.1%, 66.8% vs. 66.6%, and 66.9% vs. 63.9%, respectively, for bankrupt
versus matched firms, with no pair having a statistically significant difference (column (7)).
In contrast, when tracking the employees from one year before until three years after
bankruptcy, we find that the rates that bankrupt firm employees remain in the original firm,
industry, and county are significantly lower than those of matched firm employees. Only 32.4%
of employees stay in their bankrupt firm (vs. 43.2% in the control firm), 52.9% stay in their

11

industry of employment (vs. 61.1%), and 53.4% stay in the county in which they work prebankruptcy (vs. 63.7%). At the bottom of the table, we examine the difference-in-differences of
the attrition rates, and find that three years post-bankruptcy, employees of bankrupt firms are 14.0
(= -22.1 – (-8.2)) percentage-points less likely to stay in the firm (SE = 7.70), 8.8 percentage-points
(SE = 4.73) in the industry, and 10.5 percentage-points (SE = 4.07) in the local area, relative to the
propensity-of-staying pattern for employees of non-bankrupt firms, and to the staying rates prebankruptcy. We find qualitatively similar difference-in-differences results by comparing changes
in mobility between employees of the bankrupt firm and randomly selected workers. We also find
that employees of bankrupt firms are 7.2 percentage-points less likely (= -9.6 – (-2.4)) to remain
in LEHD relative to those of matched non-bankrupt firms, compared with their difference before
bankruptcy, although this difference-in-differences is statistically insignificant.
In summary, Table 3 shows that the employees of the bankrupt firms are comparable to
those of the matched firms along key individual characteristics observed before bankruptcy. In
contrast, the table shows that employees of bankrupt firms are significantly more likely to leave
the firm, industry, and county than those of matched firms post-bankruptcy.
[Table 3 about here]

3. Empirical Results
3.1. Effect of Corporate Bankruptcy on Employee Earnings
We employ a difference-in-differences approach to estimate earnings changes for
employees of bankrupt firms relative to earnings changes that would have occurred in the absence
of bankruptcy. Specifically, we estimate the following regression equation:
𝑦 =𝛼 +𝛾

+ 𝛽𝑥 + ∑

𝜆 𝑑[𝑘] + ∑

𝛿 𝑑[𝑘] × 𝐵𝑅 + 𝜀 ,

(1)

where i, j, c, and t index workers, industries, counties, and years, and yit is worker i’s logarithm of
real annual earnings (adjusted using the CPI in 2001 constant dollars) in year t. For a worker who
holds multiple jobs in a given year, we use the worker’s total earnings aggregated across all her

12

jobs in that year. 14 αi and γjct represent worker and industry × county × year fixed effects. xit
includes the following time-varying worker characteristics: years of experience (defined as age –
imputed education – 6) × imputed years of education and years of experience × female indicator.
We do not include imputed education or a female indicator as stand-alone variables because they
are constant at the individual level and thus are collinear with worker fixed effects. Likewise, years
of experience is implicitly controlled for as it is collinear with worker and year fixed effects. We
do not include age because it is collinear with experience and imputed education. d[k]it is an
indicator variable equal to one if year t is |k| years before or after a bankruptcy filing, and zero
otherwise (-3 ≤ k ≤ 6). BRi is an indicator variable equal to one if worker i was an employee of a
bankrupt firm (in any of the worker’s jobs) one year prior to bankruptcy, and zero if the worker
was an employee of a control firm. εit is the random error. Standard errors are double clustered at
the firm and worker level.
We use equation (1) to study employees that work at the bankrupt (or control) firms one
year prior to the bankruptcy filing (or year of matching) and analyze these employees from t-4 to
t+6. As the event indicator variables begin in t-3, the implicit benchmark is earnings in year t-4 for
both groups. The estimates of δk capture the change (from the benchmark year t-4) in earnings of
bankrupt firms’ employees in each event year relative to that of the control group. We include
worker fixed effects to control for unobserved individual-level heterogeneity. We include original
two-digit SIC industry × county × year fixed effects to control for unobserved time-varying
economic shocks specific to industry by county of the worker’s pre-bankruptcy job. Thus, our
identification of the effect of corporate bankruptcy on employee earnings is based on the variation
in earnings between employees of bankrupt and non-bankrupt firms in the same original industrycounty cell in the same year.
Although we match non-bankrupt firms on observables and include them as a
counterfactual group, and employ detailed fixed effects to control for unobserved heterogeneities
across industries and local labor markets over time, we note that our empirical approach might not
14

In the case of multiple jobs, industry and county fixed effects are based on the industry and location for her job with
highest earnings.

13

provide perfect counterfactuals. For example, a negative technology (or general economic) shock
might hit the bankrupt firm and its employees but not the matched firm and employees, even within
the same industry and local labor market. If such a shock leads to both a bankruptcy filing and a
reduction in the value of employee skills, the future earnings of the bankrupt firm’s employees
could decline even in the absence of bankruptcy filing itself. However, if economic shocks are
largely common within industries and local labor markets, our approach is likely to provide a valid
counterfactual.
3.1.1. Baseline Results
Table 4 presents the regression results, with column (5) reporting the result from the main
specification in equation (1) and other columns controlling for various layers of fixed effects. The
control group includes employees of the propensity-score matched firms in Panel A, and randomly
selected employees of non-bankrupt public firms in Panel B. The interaction terms d[k] × BR (0
≤ k ≤ 6) are the variables of interest and the coefficient estimates show that relative to the control
group, employees of bankrupt firms experience significant reductions in earnings over the years
after a bankruptcy filing. Across all specifications, most of the coefficients on the interaction terms
are significantly negative at conventional levels from year t to t+2 or t+3.
[Table 4 about here]
We examine the economic magnitude of the employee earnings losses associated with
bankruptcy filings using estimates in column (5) of Panel A, which includes worker and industry
× county × year fixed effects. We use column (5) as the preferred specification to ensure that our
estimates do not simply reflect time-varying industry-by-local labor market conditions. For
example, bankrupt firm employees may be in industries and geographic areas that experience
different economic shocks than those of control firm employees, which would affect their
subsequent earnings even in the absence of bankruptcy filings.
In column (5), the -0.128 coefficient on d[0] × BR, which is the difference between log
earnings in year t and log earnings in benchmark year t-4 for bankrupt firms, relative to the same
difference for matched control firms, implies a 12.0% decline (= exp(-0.128) – 1) in earnings in
year t relative to the counterfactual earnings. Similarly, the coefficients on d[k] × BR (1 ≤ k ≤ 6)

14

indicate an average annual earnings loss of around 10% relative to the pre-bankruptcy earnings
and to the non-bankrupt firm employees. The last few rows in the panel show that the present value
(PV) of earnings losses during the seven years since bankruptcy filing, computed using a real
discount rate of -7.4%, is 89.5% of the pre-bankruptcy annual earnings (measured in benchmark
year t-4).15 As explained in the preceding footnote, the discount rate is negative because it accounts
for workers’ uninsurable idiosyncratic risk. Alternatively, the PV of earnings losses calculated
using a real risk-free rate of 3.34% is 66.0% of the pre-bankruptcy earnings.
Figure 1 visually presents the earnings dynamics based on the coefficient estimates in
column (5) of Panel A. The figure shows that earnings of bankrupt firm employees (relative to
those of matched control firm employees) are flat during the pre-bankruptcy period from t-3 to t1. Earnings are significantly lower in the year of the bankruptcy filing and remain low from that
year onward. From t to t+6, Panel A shows that employees lose about 10% of the pre-bankruptcy
(t-4) annual earnings on average, relative to the earnings of the matched firm employees in the
respective year. Panel B presents the cumulative employee earnings losses through t+6.
[Figure 1 about here]
The other columns in Table 4 explore whether our results are robust with respect to
econometric specifications with different fixed effects. All specifications control for worker fixed
effects. In general, when we use less refined layers of fixed effects (e.g., county × year or industry
× year rather than industry × county × year), heterogeneity that we do not control for (such as
industry-by-region-specific shocks) likely biases estimates of earnings losses for bankrupt firm
employees. In column (1), which includes year fixed effects, we find that the PV of earnings losses
is 58.9% of pre-bankruptcy earnings. Columns (2) through (4) include industry × year, county ×
year, and both industry × year and county × year fixed effects, respectively. The PV’s of earnings
15

To calculate the PV of earnings losses, we use an inflation- and risk-adjusted discount rate that is appropriate given
workers’ exposure to uninsurable idiosyncratic risk. To this end, we first simulate the distribution of wealth for a
representative worker who earns average earnings in our sample and holds average financial assets in the Survey of
Consumer Finance. We then use a certainty equivalence approach to find the discount rate for this representative
worker. See Online Appendix C for details of the approach. To bookend our estimate, we also perform the PV
calculation using a real risk-free rate of 3.34%. The PV of earnings losses calculated using the real risk-free rate
ignores the worker’s (idiosyncratic) risk exposure.

15

losses in these specifications are generally smaller than that in column (5) with the PV in column
(4) closest in magnitude (68.7%).
3.1.2. Additional Results
This section examines the robustness of the baseline results using an alternative control group, by
accounting for unemployment insurance benefits in the earnings loss estimates, and by imputing
missing earnings in LEHD.
Random Sample of Workers as Control Group: Panel B of Table 4 reports the regression results
based on the randomly selected control group of workers. Comparing to the results in Table 4
Panel A, the magnitudes of earnings loss estimates are generally larger when these randomly
selected workers are used as the control group and industry × county × year fixed effects are not
employed. For example, column (2) in Panels A and B shows present value earnings losses are 40%
and 121% of the benchmark annual earnings, respectively. This result suggests that the latter
estimates may be biased upward in absolute value, likely due to lack of controls for firm-level
characteristics, particularly economic performance proxies that we use as matching variables (such
as the change in market value of assets, ROA, and market-to-book). However, we find similar
magnitudes of PV losses in column (5) in both panels, indicating that industry × county × year
fixed effects control for important unobserved heterogeneity between the bankruptcy and control
groups that may be captured by firm characteristics employed in matching. The similar column (5)
magnitudes between the two panels provide confidence in our estimates of the PV earnings loss
due to bankruptcy of about 79-89%.
Unemployment Insurance Benefits: We incorporate unemployment insurance (UI) into our
estimates because UI benefits would partially offset workers’ lost earnings after bankruptcy (see,
e.g., Hsu, Matsa, and Melzer, 2018). We collect information on regular and extended UI benefits
and estimate each eligible unemployed worker’s quarterly UI benefits, which are added to their
quarterly earnings from LEHD to get total, UI-adjusted earnings (Online Appendix D provides
details on the UI benefits calculations).16 In the baseline analysis, we define unemployment spells

16

We thank David Matsa for generously providing data on state UI benefits for this analysis.

16

as lasting a maximum of two quarters of consecutive zero or missing quarterly earnings, and
consider three or more quarters of consecutive zero or missing earnings as missing values. We reestimate equation (1) with the dependent variable being the logarithm of UI-adjusted real annual
earnings.
Panel C of Table 4 shows that the estimates for the PV of earnings losses with the UI
adjustments are generally somewhat smaller than the PV estimates without the UI adjustments.
For example, using the preferred specification in column (5), the PV’s with and without the UI
adjustments are 89.5% and 87.1% of the pre-bankruptcy earnings, respectively. Overall, the results
suggest that even after accounting for UI benefits, employees of bankrupt firms experience
significant reductions in earnings in the years following a corporate bankruptcy filing. UI has
quantitatively small effects on earnings losses in the baseline sample because the sample excludes
many potential unemployment spells that last longer than two quarters in situations in which the
worker disappears from the LEHD data.
Imputation of Missing Earnings and UI Benefits: As mentioned in Section 2.1.1, missing
earnings in the LEHD data could be due to an individual moving to a non-LEHD state, becoming
self-employed, or unemployed. To examine the implications of missing earnings for earnings loss
estimates, we provide additional estimates using two imputation approaches. 17 First, for both
treatment and control groups, we replace missing earnings with the first percentile value of the
earnings distribution in our sample. 18 This approach essentially assumes that individuals who
disappear from the LEHD data are unemployed. Alternatively, we replace missing quarterly
earnings with the maximum quarterly earnings in the last year the individual appeared in the LEHD
data; this approach effectively assumes that those who disappear move to work in non-LEHD
17

See e.g., Walker (2013), who uses similar imputation approaches to address limitations of the LEHD data. Our
approach differs from Walker’s in that he uses the average dollar earnings for cohorts as the dependent variable and
replaces missing earnings with zeros. Our dependent variable is log(earnings), and thus we use the 1 st percentile value
in lieu of earnings when workers with missing earnings records are assumed to be unemployed.
18
Specifically, when the entire annual earnings records are missing, we replace them with the first percentile value of
the annual earnings distribution. When earnings are missing for less than four quarters, we impute the missing
quarterly earnings records with the first percentile value of the annual earnings divided by four. We also impute
missing earnings with the 0.5% and 5% values of the earnings distribution and obtain similar results (see Online
Appendix Table A2).

17

states or become self-employed and earn the same wages as before. 19 , 20 Using the preferred
specification as in column (5) in Panel A, we find in Online Appendix Table A2 that when we
replace missing earnings with the first percentile, the present value of earnings losses is 158%
(114% with UI adjustments), whereas when we replace missing quarterly earnings with the
maximum quarterly earnings in the last observed annual record, the PV is 64% (62% with UI
adjustments). This considerable effect of UI on PV earnings losses in the first percentile-imputed
sample appears reasonable given that many more missing quarterly earnings are assumed to be
unemployment in this sample (where workers who disappear from the LEHD are assumed to be
unemployed until the end of the event window or sample period) relative to the baseline. In
addition, the average unemployment spell length is about eight quarters in this sample. Therefore,
the unemployed typically collect UI income for about 25% (= two /eight quarters) of the length of
unemployment spells.
3.2. Worker Reallocation as a Mechanism of Earnings Losses after Bankruptcy
Our main analysis documents a significant decline in earnings for average employees of
bankrupt firms. While informative, the baseline estimates could mask rich heterogeneity in labor
market adjustments post-bankruptcy. In this section, we condition our estimation of earnings
dynamics on whether workers reallocate across firms, industries, and geographical areas postbankruptcy, as well as on labor market characteristics and bankruptcy outcomes. The analysis
sheds light on two mechanisms through which workers may suffer earnings losses: costs associated

19

If a worker leaves an employer in an LEHD state for a non-LEHD state or becomes unemployed in the middle of a
year (e.g., June) or a quarter (e.g., February for first quarter), last annual or last quarterly earnings in her previous job
may underestimate her “true” wages in that job. To mitigate this potential bias, we impute missing quarterly earnings
using the maximum of the quarterly earnings in her last observed year. This approach is of little economic consequence
because if we were to instead impute using the last observed annual earnings divided by four, the PV earnings losses
are of similar magnitude (see Online Appendix Table A2).
20
By analyzing real earnings throughout the paper, we avoid the effects of inflationary earnings growth. Nonetheless,
real earnings may grow and the growth rate may differ conditional on the worker’s characteristics. We calculate the
average growth rate of real annual earnings by gender, education, and experience groups using the LEHD data (we
use a 10% random sample to reduce computational burden). We then apply these growth rates of real earnings to
imputed earnings in columns (5) and (6) in Online Appendix Table A2, where missing quarterly earnings are imputed
at the maximum quarterly earnings in the last observed annual record or the last observed annual earnings divided by
four.

18

with worker mobility and loss of firm- or industry-specific human capital. We consider the results
in this section suggestive and thus interpret them with caution, given that worker selection on
observable and unobservable characteristics (e.g., ability) is likely to affect both the conditioning
variable (e.g., whether switching firms or employed in smaller local labor markets) and postbankruptcy earnings. We partly mitigate this concern by including worker fixed effects in the
empirical specification. In this and other sections throughout the rest of this paper, we use the
specification that includes worker and two-digit SIC industry × county × year fixed effects and use
the matched control group (see column (5) of Table 4, Panel A, and related discussion in Section
3.1).
3.2.1. Worker Movement and Post-Bankruptcy Earnings
If an important part of earnings losses during labor market adjustments is due to the loss
of human capital specific to firms (Becker, 1962) or industries (Neal, 1995), then the costs of
bankruptcy might be largely borne by workers who leave their firms or industries. To test these
predictions, in the following regression specification we employ an indicator variable Stay, which
is equal to one if a worker who was in a given firm, two-digit SIC industry, or county in year t-1
stays in the same firm, industry, or county through year t+3:
𝑦 =𝛼 +𝛾

+ 𝛽𝑥 + ∑

𝜆 𝑑[𝑘] + ∑

𝑑[𝑘] × 𝐵𝑅 × (𝛿 𝑆𝑡𝑎𝑦 + 𝜃 (1 − 𝑆𝑡𝑎𝑦 )) + 𝜀 , (2)

where the coefficients δk represent the effect of bankruptcy on earnings of the bankrupt firm
employees who stay in the firm (or industry, or county), and the coefficients θk represent the effect
on earnings of the bankrupt firm employees who do not stay. 21 We stratify the treatment (i.e.,
bankruptcy) group based on whether workers stay with their firms (or industries, or counties) until
year t+3. We do not stratify the control group, so our analysis decomposes earnings losses of the
treated group relative to the average worker in the control group (see e.g., Walker (2013) for a
similar stratification).
[Table 5 about here]
21

Note that the specification in equation (2) does not include the stand-alone Stay indicator because it is collinear with
worker fixed effects. Also, we find similar results if we alternatively define the Stay indicator equal to one if a worker
stays in the same firm, two-digit SIC industry, or county through year t+5 (see Online Appendix Table A3).

19

We first examine whether the magnitudes of earnings losses are different between the
employees who stay with the bankrupt firm and those who leave the firm (Table 5, Panel A).
Comparing the estimates indicates that “firm leavers” (column (2)) fare significantly worse than
“firm stayers” (column (1)) during the years after the bankruptcy filing. Using the present-value
approach described in Section 3.1, workers who leave the firm lose on average 120% of their
annual earnings for the seven years from t to t+6 (significant at the 1% level), while those who
stay with the firm experience a statistically insignificant gain of 38%. 22
We next unpack the earnings loss of the workers who leave the bankrupt firm along another
dimension: whether they leave the industry and/or county. If an employee’s human capital (e.g.,
skill set) is specific to her original industry, the earnings loss would be more pronounced for
“industry leavers,” all else equal. Consistent with this prediction, columns (5) and (6) show that
among county leavers, earnings losses for industry leavers are substantially greater than for those
for industry stayers, with the difference significant at the 1% level. Similarly, columns (3) and (4)
examine county stayers and also find that the industry leaver effect (= 19.0% – 79.1%) is negative
and significant at the 5% level. These results imply that the cost of corporate bankruptcy is borne
by workers who leave their industries (and not by those who stay in the same industries even
though leaving their original firms), consistent with industry-specific human capital playing an
important role in labor market adjustments after bankruptcy. 23 As noted above, however, the
22

Note that the above results show the earnings losses of stayers and leavers in the treated group relative to the average
worker in the control group. Instead of using the average worker in the control group as the benchmark, in a separate
specification we examine whether stayers (leavers) of the bankrupt firm fare worse than the stayers (leavers) of the
control firm by including the Stay indicator (and 1 – Stay indicator) uninteracted with a BR indicator but interacted
with the event time indicators in equation (2). That is, we replace ∑
𝜆 𝑑[𝑘] in equation (2) with
∑
𝑑[𝑘] × (𝜇 𝑆𝑡𝑎𝑦 + 𝜋 (1 − 𝑆𝑡𝑎𝑦 )). The results for this conditional analysis that stratifies both the bankruptcy
and non-bankruptcy groups (in Online Appendix Table A4) show that both stayers and leavers of bankrupt firms lose
significant earnings relative to their counterparts in non-bankrupt firms. In particular, the finding that stayers of
bankrupt firms also lose considerable earnings relative to stayers of non-bankrupt firms suggests that earnings losses
associated with corporate bankruptcy may be beyond the losses due to just job displacement. However, we interpret
this implication with caution given the different selection for stayers (or leavers) of bankrupt and non-bankrupt firms.
23
Kambourov and Manovskii (2009) show that once controlling for occupation-specific tenure (based on the Standard
Occupational Classification), tenure in an industry (based on SIC) has a limited effect on wages, suggesting that the
effect of industry switches on earnings that we find may be (in part) due to occupation switches (which are likely
correlated with industry switches). Unfortunately, we are unable to empirically examine this issue because the LEHD
data do not contain information on the occupation of employment.

20

association between switching industry and earnings losses could be due to selection. For example,
it is possible that more able workers stay in the industry while worse workers leave, leading to the
differential earnings patterns across the groups.
In terms of county stayer status, the comparison of columns (3) and (5) and columns (4)
and (6) shows that whether workers remain in the same industry or not, switching counties is
associated with significantly lower post-bankruptcy earnings at the 1% level. This considerable
difference appears to reflect a more onerous transition period for those who relocate across
geographical areas, or due to selection (on ability, for example) in which less able workers “have
to” move outside the original county post-bankruptcy.
In sum, while the analysis in this subsection does not allow us to make a statement about
causality of earnings losses post-bankruptcy, it sheds light on potential drivers of the losses, such
as industry-specific skills and costly worker movements across industries and geographies.
3.2.2. Heterogeneity Conditional on Firm and Labor Market Size and Bankruptcy Outcomes
The analysis in the previous section suggests that costs to worker mobility and loss of
specific human capital are potential mechanisms through which workers suffer earnings losses. To
explore these mechanisms further, in this section we examine employees’ post-bankruptcy
earnings conditional on labor market characteristics and bankruptcy outcomes using a modified
version of equation (2). Specifically, for this analysis we stratify both the treatment (bankruptcy)
and control groups of employees based on the conditioning variables, which include firm size,
local labor market size, and bankruptcy liquidation outcome. Then we compare the present value
of earnings losses across the groups in Table 5, Panel B.
We first explore the effect of bankruptcy on earnings conditional on the size of firms
measured by book assets. Workers in larger firms may fare better after bankruptcy than those in
smaller firms because of better opportunities to reallocate within (e.g., to another division in the
internal labor market) or across firms (Lazear and Oyer, 2004; Tate and Yang, 2015) or because
of better opportunities to develop general skills that improve their job mobility. Moreover, Graham
(2022) shows that concerns about leverage-induced financial distress are greater in small firms.
Consistent with these notions, the estimates in columns (1) and (2) in Panel B show that the present

21

value of earning losses for employees of smaller firms is 99% of pre-bankruptcy annual earnings
and statistically significant, whereas it is 62% for employees of larger firms (defined as book assets
larger than the 75th percentile) though this difference is not significant.
Second, we explore whether the impact of bankruptcy on earnings is related to the size of
labor markets. The analysis is motivated by the extant literature arguing that large labor markets
reduce search frictions and thus make it easier for workers to find jobs requiring similar skills
(Diamond, 1982; Moretti, 2011). Following the labor market search literature (e.g., Petrongolo
and Pissaride, 2006; Manning and Petrongolo, 2017; Kim, 2020), we measure the size of the labor
market at the local level. In particular, we sort industry-county cells at the 75 th percentile of the
distribution of total employment (from the LBD) into “large” vs. “small” (or “thick” vs. “thin”)
labor markets. 24 The results, reported in columns (3) and (4) in Table 5, Panel B, show that
earnings decline less in larger labor markets after bankruptcy, plausibly because workers in these
markets can more easily find jobs that match their skill set. Specifically, the present value of
earning losses for employees facing a smaller labor market is 116% and statistically significant,
while it is 24% for employees of firms facing a larger labor market and statistically insignificant.
Combined with the results in Panel A, these results highlight the importance of employment
opportunities in the same industry (that presumably require similar human capital) and local area
(that reduce search costs) in mitigating the adverse effects of bankruptcy on earnings. We note that
the relation between employer and labor market characteristics and earnings losses postbankruptcy could be due to selection, and thus we interpret these results as an association instead
of a causal relation.
Lastly, in columns (5) and (6) we estimate the earnings patterns of employees for ultimately
liquidated versus non-liquidated bankrupt firms. To the extent that workers are more likely to leave
the firm after a bankruptcy that eventually ends in liquidation, earnings losses could be greater in
liquidation. In addition, liquidation may indicate poor economic or labor market conditions, under
which worker reallocation is costlier, leading to greater earnings losses. We find that the PV

24

We sort the sample at the 75th percentile, given that the distribution of these variables is right-skewed.

22

earnings losses may be slightly larger in absolute magnitude for liquidated firms than for nonliquidated firms. Columns (5) and (6) show that the PV of earnings losses amounts to 97% (of prebankruptcy annual earnings) for liquidated firms and 84% for non-liquidated firms. The
insignificant difference between liquidating and non-liquidating bankrupt firms appears to be due
to different timing of earnings losses across the groups – employees of liquidated firms experience
larger losses right after bankruptcy (when many workers likely leave the firms) but smaller losses
in later years, when their transition to next jobs is likely complete; in comparison, employees of
non-liquidating firms experience considerable losses throughout the seven years post-bankruptcy
(see Online Appendix Figure A2). The finding suggests that while liquidation may accelerate
earnings losses, employees ultimately experience significant earnings losses whether the bankrupt
firm liquidates or not.
3.3. Regression Discontinuity Analysis of Loan Covenant Violations
Our main analysis thus far examines earnings dynamics for employees of bankrupt and
matched firms and provides evidence consistent with ex ante comparable workers experiencing
substantial earnings losses due to corporate financial distress. However, the analysis based on
matching on observables and detailed fixed effects may not completely rule out alternative
explanations concerning unobserved economic shocks affecting the different earnings trajectories
of the treated and control groups. In this section, we provide complementary evidence for a causal
effect of financial distress on employee earnings by using a regression discontinuity design (RDD)
around loan covenant violations. Specifically, we follow Falato and Liang (2016) to construct a
sample of firms in the DealScan and Compustat databases, match them with employees in LEHD
from 1995 through 2008, and estimate the following equation: 25, 26
∆𝑦

→

=𝛼 +𝛾

+ 𝐵𝑖𝑛𝑑

+ 𝛿𝑧

+ 𝛽𝑥

+𝜀

,

(3)

where ∆yift-1→t is the first difference in worker i’s (who is employed by firm f in year t-1) logarithm
of real annual earnings in year t relative to year t-1, αf and γjct represent firm and two-digit SIC

25
26

We thank Antonio Falato for sharing DealScan-Compustat merged data and related programs.
See also Chava and Roberts (2008) for an RDD analysis of loan covenant violations on corporate investment.

23

industry × county × year fixed effects; Bindift is an indicator variable equal to one if worker i is an
employee of firm f that violates a loan covenant concerning the current ratio or net worth in year
t, and zero otherwise; zift includes the relative distances from covenant violation thresholds for the
current ratio (CR) and net worth (NW) and their second-order polynomial terms; and xit-1 includes
the following worker characteristics: imputed education, female indicator, imputed education ×
female indicator, years of experience, imputed education × years of experience, and female
indicator × years of experience. εift is the random error. Standard errors are double clustered at the
firm and worker level. Following Falato and Liang (2016), we use a sample of firms within +/20% of the relative distance to covenant violation thresholds (“discontinuity sample”). The sample
excludes financials, firms with employment less than 10, and those incorporated outside the U.S.
In addition, we include up to second-order (i.e., square) terms of the relative distances (see e.g.,
Gelman and Imbens, 2019).
First, we verify that there is no “jump” in worker characteristics (e.g., annual real earnings,
imputed education) around the violation threshold before potential violation, consistent with
employees of firms “just above” and “just below” the threshold being comparable ex ante (see
Online Appendix Table A6). In addition, Figure A3 in the Online Appendix shows that there is a
sharp break in the average earnings growth rate of employees of firms that fall just below the
covenant violation threshold relative to firms that fall just above the threshold. Second, we present
the results of estimating equation (3) in Table 6. The coefficients on Bind show that a loan covenant
violation leads to a 3.6-5.5 percentage-point (e.g., exp(-0.037) – 1 = -0.036 from column (1))
decline in the worker’s earnings growth rate, which is significant at the 5% level across the
specifications. We obtain similar results when we also include second-order terms of relative NW
and CR in column (2). The results are robust when we include firm fixed effects (columns (3) and
(4)), which control for unobserved firm-level heterogeneity in earnings growth due, e.g., to
differences in productivity growth.
[Table 6 about here]
To the extent that a covenant violation represents a form of financial distress, potentially a
“first step” towards bankruptcy, the results for declining earnings growth due to covenant

24

violations are qualitatively consistent with the existence of a causal effect of financial distress on
employee earnings losses, which we analyze in the main analysis using corporate bankruptcy
events. In this sense, the analysis of loan covenant violations complements the main analysis and
supports the interpretation that the employee costs of financial distress could affect firms’ capital
structure decisions ex ante. This is the topic we examine in the next section.

4. Wage Premia for Bankruptcy Risk and Capital Structure Implications
In competitive labor markets, employees exposed to higher risk of earnings loss (due to,
for example, unemployment or moving to lower-paying jobs) would plausibly demand an ex ante
wage premium to compensate for the risk (Abowd and Ashenfelter, 1981; Topel, 1984). 27 In the
context of bankruptcy risk, such a wage premium represents an ex ante cost of bankruptcy for the
firm, creating a disincentive to use debt. This section gauges whether this indirect cost of
bankruptcy is of sufficient magnitude to affect corporate debt choices. In Section 4.1 we estimate
the effect of leverage on wage premia by regressing wages on leverage using the LEHD data.
Given that regression estimates may be affected by omitted variables, in Section 4.2 we
alternatively derive wage premia using a binomial valuation model. In Section 4.3, we show that
implied wage premia from the wage regression and from the valuation model have similar orders
of economic magnitude. Finally, in Section 4.4 we compare our implied wage premium to the tax
benefit of debt to explore capital structure implications. 28
4.1. Estimating Wage Premia Using Wage Regressions

27

See e.g., Rosen (1986) for theoretical underpinnings of compensating (or equalizing) differences. Agrawal and
Matsa (2013) point out that even if workers do not gauge their employment stability by observing direct signals of the
firm’s financial conditions such as financial leverage and credit ratings, they can rely on indirect signals from
management, the media, and other aspects of economic conditions. Consistent with workers’ ability to understand the
financial status of firms, Brown and Matsa (2016) find that job seekers accurately perceive firms’ financial health and
act upon it by reducing labor supply to distressed firms.
28
Our approach uses compensating wage differentials to directly estimate the cost of financial distress resulting from
the employee earnings losses due to bankruptcy. See Brown and Matsa (2016) and Baghai, Silva, Thell, and Vig (2021)
for evidence that highly levered firms lose high-quality job candidates and employees due to poor job stability. Brown
and Matsa (2016) also provide an empirical approach to estimate the indirect cost of financial distress due to
deterioration of worker quality in financial distress.

25

We first estimate wage premia for the risk of bankruptcy using a wage regression approach
on a sample of workers with ages between 22 and 55 who move across jobs at public firms in
LEHD. In particular, we regress movers’ highest quarterly earnings in the last four quarters right
before and in the first the four quarters right after the movement on the leverages of their old and
new employers, respectively. This choice of highest quarterly earnings as a measure of wages is
in recognition that quarterly earnings immediately before or after a move likely under-estimate the
“true” full-quarter wage rates for the worker because of the “move-within-the-quarter” issue.
We focus on earnings of “new hires” given that firms are hesitant to adjust wages for
incumbent workers (e.g., Dickens et al., 2007), likely due to concerns about fairness, adverse
selection, and turnover (Blinder and Choi, 1990; Campbell and Kamlani, 1997). In addition,
Brown and Matsa (2016) find evidence that job changers accurately perceive firms’ financial
health and firms in distress indeed offer higher posted wages, suggesting that wages of job
changers are likely sensitive to the company’s leverage. Sample firms are in Compustat from 1986
to 2008, and have S&P credit ratings between AA+ and B- or unrated. 29 We apply the same criteria
for book assets as in the bankrupt firm sample used in Section 2. 30 We drop spurious moves such
as changes in firm identifiers due to apparent mergers and acquisitions from the sample. This
procedure yields about 9.1 million quarterly earnings observations of job movers before and after
their movements.
Using this sample, we estimate a standard earnings equation augmented by market leverage
of the employer, a proxy for the risk of bankruptcy, as follows:
𝑦

=𝛼 +𝛼 +𝛼

+ 𝛽𝑀𝑎𝑟𝑘𝑒𝑡 𝑙𝑒𝑣𝑒𝑟𝑎𝑔𝑒

+ 𝛾𝑥 + 𝛿𝑧

29

+𝜀

,

(4)

We exclude firms with AAA ratings given that they are rare in practice. We also exclude firms with ratings below
B- in an attempt to focus on the ex-ante wage premium outside of ex-post financial distress. For example, Benmelech,
Bergman, and Enriquez (2012) show that employees of financially distressed firms (especially those with weak
bargaining power) experience wage cuts, which would induce a negative bias into our estimate of the wage-leverage
relation. Similarly, Almeida and Philippon (2007) stop at a B rating in their analysis. We present the results when all
these firms are included in the sample (in Online Appendix Table A7, Panel A). The table shows very similar results
for the effect of leverage on wages.
30
That is, we require that book assets be greater than $100 million in 1980 constant dollars and less than or equal to
the assets of the largest firm in our bankrupt and matched firm sample.

26

where yifjct is the logarithm of quarterly real earnings (in 2001 constant dollars) for worker i
employed in firm f, industry j, county c, and year t; αi, αf, and αjct represent worker, firm, and twodigit SIC industry × county × year fixed effects; and Market leverageift is the market leverage ratio
of firm f (where worker i is employed) in year t. xit is a set of worker-level control variables
including a female indicator × years of experience and imputed education × years of experience.
zift includes the following firm-level control variables: log book assets, market-to-book, ROA, and
asset tangibility for firm f that employs worker i in year t. εift is the random error. Standard errors
are double clustered at the firm and worker level. Controlling for employee ability is important for
identifying the effect of financial distress risk on wages (Berk, Stanton, and Zechner, 2010) and
compensating differentials in general (Brown, 1980). Therefore, we control for both worker and
firm fixed effects (e.g., Abowd, Kramarz, and Margolis, 1999; He and Maire, 2020) as well as the
observed worker and firm characteristics.
[Table 7 about here]
Consistent with the existence of compensating differentials for corporate bankruptcy risk,
Table 7 shows evidence that employees of firms with higher leverage are paid higher wages, all
else equal. The coefficient on market leverage in column (1) suggests that a ten-percentage-point
increase in leverage is associated with a 0.59% (= 10% × (exp(0.057) – 1)) increase in wages
(significant at the 1% level). Importantly, this result is based on a sample of public firm employees
from the LEHD data. Thus, our estimates refine those in Chemmanur, Cheng, and Zhang (2013),
which are based on firm-level aggregate wage bills and executive compensation data. Also, our
results complement Brown and Matsa (2016), who show that firms with higher risk of financial
distress (proxied by CDS prices) offer higher wages to job applicants.
We next examine the heterogeneity in wage premia for employees facing different risks of
earnings loss. We split the full sample at the 75th percentiles of firm size (measured by book assets)
and local labor market size (measured by employment in a given two-digit SIC industry and county
cell) distributions.31 Columns (2) through (5) show that significant wage premia exist only for
31

As in Section 3.2.2, we sort the sample at the 75th percentile, given that the distribution of these variables is rightskewed.

27

workers employed by smaller firms and in smaller local labor markets. The difference in wage
premium estimates between workers in smaller versus larger firms (larger labor markets) is
significant at the 10% (5%) level. A concurrent paper by Dore and Zarutskie (2018) also estimates
earnings equations using LEHD data and finds that an increase in book leverage is associated with
higher wage growth, particularly in smaller (relative to larger) local labor markets. Combined with
our finding in Table 5, Panel B that bankruptcy is associated with greater earnings losses for
workers employed by smaller firms and in smaller local labor markets, the results in this section
suggest that workers are paid a significant wage premium for bankruptcy risk, especially for job
movers.32
We further sort the sample of bankrupt and matched firm employees at the 75 th percentile of
ages in year t-1. We find that older workers (age > 75th percentile) lose 108% of their prebankruptcy annual earnings over the seven years after bankruptcy, whereas younger workers lose
79%. However, we find that older workers do not receive a larger magnitude of financial leverage
wage premium than do younger workers – a ten-percentage-point increase in market leverage is
associated with 0.48% and 0.64% increases in wages for older and younger workers, respectively.
That is, unlike our tests on firm size and labor market size above, the PV (earnings losses) and

32

We report additional results for the sample including (a 30% random sample of) all workers (movers and nonmovers) in Online Appendix Table A7, Panel B. The results show that leverage is significantly positively associated
with wages in regressions without firm fixed effects, while the relation becomes insignificant when controlling for
firm fixed effects. The results are consistent with our earlier discussion of the wage rigidity of incumbent workers,
which implies that it may be difficult to detect the relation between leverage and wages using within-firm variation.
It would be particularly difficult for the firm to reduce incumbent workers’ wages when it decreases leverage due to
downward wage rigidity. Even when a firm increases its leverage, it might not increase wages immediately because
labor market frictions (such as specific human capital and search costs) may prevent incumbent workers from moving
quickly to other firms with a lower leverage and/or higher wages. Intuitively, the firm would only need to increase
wages for higher leverage when the worker is willing to switch jobs given the associated mobility costs (See Manning
(2003, chapter 8.2) for a general discussion of how worker mobility costs, i.e., imperfect competition in labor markets,
render workers not fully compensated for job disamenities like risk of earnings losses). Therefore, for a ‘modest’
increase in leverage, we might not observe an immediate adjustment in the wages of incumbent workers. In contrast,
for job movers, who by revealed preference make job-to-job moves reflecting their preferences, wages are more likely
to promptly reflect a premium (discount) for higher (lower) leverage. As such, there could be a difference in the
promptness between changes in the wages of incumbent workers and job movers, and the process for wages of all
employees to fully incorporate a premium or discount for leverage changes could be slow. Under these conditions, the
leverage-wage relationship may be easier to detect as a cross-sectional phenomenon that shows up in regressions
without firm fixed effects and/or through the mover sample.

28

wage premium results do not align for worker age. A possible explanation for why older workers
appear not to receive a larger wage premium conditional on leverage is that, while older workers
face greater earnings losses upon bankruptcy, they also have less labor market mobility, which
could imply that the labor market they face is “less competitive.” As labor markets become less
competitive, firms are less likely to pay compensating differentials.
Although we control for firm, worker, and industry × county × year fixed effects, as well as
firm and worker characteristics in our wage equation, one caveat of the analysis is that the wage
premium estimates may be biased due to omitted variables. One such potential omitted variable is
the quality of corporate governance. For example, Cronqvist et al. (2009) find that entrenched
CEOs tend to pay higher wages to workers, whereas Berger, Ofek and Yermack (1997) find
evidence that entrenched CEOs seek to avoid debt usage, which implies that the quality of
corporate governance might be negatively (positively) correlated with wages (leverage). Thus,
omitting variables capturing the quality of corporate governance, which likely varies over time
even within firms, could lead to under-estimating the effect of leverage on wages.
Similarly, managerial characteristics such as overconfidence could bias our estimated
relation. CEO overconfidence is positively associated with the firm’s leverage (Malmendier et al.,
2011; Ben-David et al., 2013). In addition, it is conceivable that overconfident managers may
overpay workers because they believe their firms are more valuable than the market value.
Therefore, omitting a control for CEO overconfidence (which varies within firms over time) could
lead to over-estimating the relationship between leverage and employee wages. We thus employ
an alternative approach to quantify wage premia in the next section.
4.2. Implied Wage Premia Using a Binomial Valuation Model
Our alternative approach to derive wage premia builds on the notion of compensating
differentials. In a competitive labor market, firms with different risks of bankruptcy pay risk-averse
employees the same risk-adjusted present value of expected wages (or provide the same level of
expected utility), all else equal. In particular, the contracted wages for a firm with a higher
probability of bankruptcy should be higher due to greater expected personal costs of bankruptcy
for the firm’s employees. This argument implies that the wage premium that a worker demands

29

should equal the expected PV of the earnings losses due to the likelihood of her employer filling
bankruptcy in that year.
Online Appendix A develops a framework to estimate the wage premia demanded by an
employee working for a risky firm. The appendix shows that a firm’s wage premium relative to
wages at a risk-free firm is equal to

(

)

, where rf is the risk-free rate, q is the annual risk-adjusted

(i.e., risk-neutral) bankruptcy probability, and

is the present value of employee net earnings

losses (with UI adjustments) given bankruptcy as a fraction of pre-bankruptcy wages. Intuitively,
at the beginning of a year, a worker employed by a risky firm with an annual bankruptcy
probability q will face an expected earnings loss of q×(NL/W) by the end of the year, for which
the worker demands a risk premium equal to its present value

(

)

. We focus on “net” earnings

losses given that part of the present value of earnings losses might be due to the loss of a
compensation differential that existed pre-bankruptcy between a riskier bankrupt firm and a less
risky matched firm. The net earnings losses are equal to earnings loss estimates in Section 3.1
times an adjustment factor.33
Because no direct estimate of the risk-neutral probability of bankruptcy, q, is available, we
begin our estimation of q from the risk-adjusted probability of default that Almeida and Philippon
(2007) derive from bond risk premia times the conditional probability of bankruptcy given default
(Prob(Bankruptcy|Default)). Using Moody’s Default and Recovery Database (DRD) from 1981 to
2013, we find that 66.4% of public default firms ultimately file for bankruptcy. Similarly, the
Altman-Kuehne/NYU Salomon Center Bond Master Default Database shows that 58.7% of firms
in default end up filing for bankruptcy protection from 1981 to 2014. 34 Thus, we use 60% as our
estimate of Prob(Bankruptcy|Default). Given that bankruptcy is likely to occur in bad states, the

33

In Online Appendix A.3, we show that the adjustment factor is equal to

years t to t+6.
34
We thank Edward Altman for sharing his dataset of default events.

30

, where 𝑓 is the annuity factor from

conditional risk-neutral probability may be higher than the conditional natural probability. We thus
check the sensitivity of our results by increasing the conditional probability from 0.6 to 0.9 and
find that this results in the wage premium between BBB and AA-rated firms increasing by 0.33%
of firm value (from 1.99% to 2.32%; see column (2) in Table 10, Panel A), suggesting little
sensitivity of our results to the conditional probability. 35
Since Almeida and Philippon (2007) provide ten-year cumulative risk-adjusted default
probabilities, we convert the ten-year probability into a one-year probability. Specifically, in
column (2) in Table 8, for each credit rating group, we compute the annual risk-adjusted
probability of bankruptcy as q = 1 - (1-0.6 × q10)1/10, where 0.6 is the probability of bankruptcy
conditional on default discussed above, and q10 is the ten-year risk-adjusted default probability
from Almeida and Philippon (2007) shown in column (1).
[Table 8 about here]
Importantly, the risk-adjusted probability of default for bondholders is likely different from
that for workers who likely hold less diversified portfolios and thus are exposed to more
idiosyncratic risk. To account for the difference, we use a certainty equivalence approach to
estimate the ratio of annual risk-adjusted default probability for workers to that for bondholders,
using the portfolio holdings information of average bondholders and workers in the Survey of
Consumer Finance. We then use the ratio to adjust q to obtain 𝑞 , the annual risk-adjusted
bankruptcy probability for workers. Column (3) in Table 8 shows 𝑞 , which we use to calculate
the implied wage premium as a fraction of wages paid by the firm to the worker:

(

)

. Online

Appendix B provides details of the estimation approach for 𝑞 .
Column (4) in Table 8 shows that the implied wage premium accounts for a notable portion
of wages, particularly for firms with low credit ratings. The implied wage premium accounts for
0.29% of wages for AA-rated firms and 0.94% for BBB-rated firms.
4.3. Comparing Regression-Based and Implied Wage Premia

35

Given that default is also likely to happen in bad states, the “true” conditional probability might be smaller than 0.9.

31

In this section, we compare the implied wage premium derived from the binomial valuation
model in Section 4.2 to the wage premium estimated from the wage regression in Section 4.1.
Given that our wage regression does not benchmark with risk-free firms, we use AA-rated firms
as a benchmark for our regression-based wage premia estimates. As such, Table 9 compares the
wage premia that BBB-rated firms pay relative to AA-rated firms between these two approaches.
Full sample comparisons are in column (1), while columns (2) and (3) show comparisons for
subsamples based on firm assets size and columns (5) and (6) based on labor market size.
Panel A shows that using the binomial valuation model, the additional implied wage
premium due to moving from AA to BBB rating is 0.66% of wages for the full sample. We derive
the implied wage premia for subsamples in columns (2) through (7) similarly following the process
explained in Section 4.2, except that we use the UI-adjusted PV earnings loss estimates for those
subsamples from Online Appendix Table A8, Panel B, columns (1) through (4) (e.g., -97.7% of
pre-bankruptcy earnings for employees of smaller firms). As shown in Table 9, the implied wage
premia are larger in subsamples for which the post-bankruptcy earnings losses are greater (i.e., for
smaller firms and smaller labor markets). The difference between the two groups is significant for
labor market size (at the 5% level) but not for firm size.
[Table 9 about here]
Next, in Panel B we report the estimates for the differential wage premium between BBB
and AA-rated firms using the wage regression results. We begin with the coefficient on market
leverage in column (1) of Table 7 (0.057), which implies that a one-percentage-point increase in
leverage is associated with a 0.059% (= exp(0.057) – 1) increase in wages. We then calculate the
difference in market leverage between BBB and AA-rated firms in our sample, controlling for
aggregate changes in leverage using year fixed effects. This leverage difference is equal to 11.3%.
Lastly, the differential wage premium between BBB and AA-rated firms is equal to the leverage
difference (11.3%) times the coefficient on leverage in the wage regression in percentage terms
(0.059), which is 0.64%. The wage premia for subsamples are calculated similarly except that we
use coefficients on leverage for different subsamples in Table 7, columns (2) to (5). Panel B of
Table 9 reinforces the conclusion obtained from Panel A that wage premia are larger when

32

employees face potentially larger earnings losses from bankruptcy (i.e., for smaller firms and for
firms operating in smaller local labor markets).
Comparing Panels A and B of Table 9 shows that the differences in wage premia between
BBB and AA-rated firms are of similar magnitudes in the two different approaches we employ.
4.4. Implied Wage Premia and Capital Structure Implications
In this section, we build upon the implied wage premia for risky firms (relative to risk-free
firms) based on the binomial valuation model (Section 4.2) and compute the capitalized value of
implied wage premia for firms with different credit ratings. The capitalized wage premia,
interpreted as indirect costs of financial distress, are then compared with tax benefits of debt for
firms with different credit ratings. We are not able to compute the capitalized wage premia over
risk-free firms using the wage regression approach (Section 4.1) because the wage regression is
conducted on firms with credit risk. 36
4.4.1. Capitalized Value of Implied Wage Premia as Indirect Costs of Financial Distress
Given that tax benefits of debt are often measured as a fraction of firm value (e.g., Graham,
2000), we compute the capitalized value of implied wage premia as a fraction of firm value. First,
we convert the wage premium as a fraction of wages (Section 4.2),
as a fraction of market value of the firm as

(

)

=

(

)

×

(

)

, into the annual premium

, where

is the ratio of total

wage bill and market value of the average sample bankrupt firm in year t-4. Second, we compute
the present value of the wage premium for a firm with an infinite horizon, and derive that the firm’s
capitalized value of wage premia relative to a risk-free firm as a fraction of firm value is equal to
(

)

(see Section A.2 of the Online Appendix for details). 37
[Table 10 about here]

36

We benchmark against risk-free firms in calculating the PV of wage premia because the existing estimates on
expected costs of financial distress and tax benefits are relative to risk-free firms.
37
Almeida and Philippon (2007) use similar valuation models and risk-adjusted default probabilities to compute the
expected costs of financial distress (but they do not address employee costs or the wage premia that we compute).

33

Column (1) in Panel A of Table 10 shows the annual wage premium as a fraction of firm
value for firms with different credit ratings. Column (2) shows that the capitalized value of the
wage premium is a nontrivial fraction of firm value, particularly for firms with low credit ratings.
The present value of the wage premium is 1.05% of firm value for AA-rated firms, and increases
to 3.04% for BBB-rated firms. The increase in wage premia for firms with lower credit ratings
(and thus higher leverage) may potentially offset the increased tax benefits of debt for these more
highly-levered firms. In the next subsection, we explore the implications of bankruptcy-induced
wage premia for capital structure decisions.
4.4.2. Comparing Implied Wage Premia to Tax Benefits of Debt and Expected Costs of
Financial Distress
Comparing columns (2) and (3) in Panel A of Table 10 indicates that the present value of
implied wage premia is a significant fraction of the tax benefits of debt. For AA and BBB-rated
firms, tax benefits of debt equal 2.51% and 5.18% of firm value, respectively, while wage premia
are 1.05% and 3.04% of firm value. Thus, wage premia offset 42% and 59% of the tax benefits of
debt for AA and BBB-rated firms, which appears to be of large enough magnitude to affect capital
structure choices.
To further assess the importance of wage premia, we explicitly factor in non-wage costs of
financial distress by benchmarking implied wage premia with the tax benefits, net of the expected
financial distress costs estimated by Almeida and Philippon (2007) and Elkamhi, Ericsson, and
Parsons (2012). Both papers use risk-neutral probabilities to estimate expected costs of financial
distress. However, as Elkamhi et al. (2012) point out, Almeida and Philippon (2007) may overstate
the expected costs of financial distress by applying Andrade and Kaplan’s (1998) 10-23%
proportional losses to the firm’s current (i.e., potentially far outside default) value. In contrast,
Elkamhi et al. (2012) use the firm’s value near default, on which Andrade and Kaplan (1998)
derive their 10-23% estimates. As a result, the expected distress cost estimates from Elkamhi et al.
(2012) are considerably smaller than those from Almeida and Philippon (2007). For example,
Elkamhi et al. (2012) find that a BBB-rated firm’s expected costs of financial distress are only 0.5%
of firm value, which is significantly smaller than 4.53% in Almeida and Philippon (2007).

34

Based on the estimates from Elkamhi et al. (2012), when a firm’s credit rating decreases
from AA to BBB, the tax benefits of debt net of expected cost of financial distress increase by
2.17% (= 2.67% – 0.5%, last row of columns (3) and (5) in Table 10, Panel A). This is close to our
estimate of the difference in the wage premium of 1.99% between BBB and AA-rated firms
(column (2)). Therefore, our estimated wage premium appears to account for a substantial portion
of the gap between tax benefits of debt and the expected cost of financial distress documented in
the literature.
The capitalized value of implied wage premia in Panel A of Table 10 implicitly assumes that
all employees receive the premium irrespective of when they are hired, which could be interpreted
as a “frictionless” wage premium in which all employees are fully compensated for the earnings
loss risk due to financial leverage. For example, the capitalized value of implied wage premium of
a BBB-rated firm reflects the wage premium when all workers in a BBB-rated firm are
compensated for the wage premium due to the difference in bankruptcy risk between BBB and
risk-free firms. As we discussed earlier, in practice, the dynamics of a firm’s wage adjustments
depend on its wage rigidity or workers’ mobility costs. Thus, at a given point in time, a firm’s
wage premium could deviate from the “frictionless” value, with only a portion of the workers
receiving the wage premium. In Panel B of Table 10, we illustrate this point quantitatively by
presenting the implied wage premia relative to firm value assuming that only recently hired
employees are compensated for the risk of bankruptcy. The panel shows that the implied wage
premia for the newly hired employees still represent a considerable fraction of firm value, given
that the wage bill of workers with one, two, and three years of job tenure are 25%, 19%, and 14%
of the average sample bankrupt firm’s total wage bills, respectively.
We note that our estimates of wage premia are larger than those in Agrawal and Matsa (2013,
AM). AM find that, assuming no wage replacement by unemployment insurance benefits, firms
with AA and BBB ratings respectively pay 0.16% and 1.54% wage premiums. Matsa (2018) argues
that the differences between AM’s (2013) estimates and ours are understandable for several
reasons. For example, AM’s estimates account for unemployment risk only, while our estimates
also include earnings losses for employees who stay with a bankrupt firm (see Online Appendix

35

Table A4). Also, our estimates include earnings losses following an unemployment spell. These
differences help explain the somewhat larger magnitude of our wage premium estimates relative
to AM’s.
Overall, our estimates for wage premia associated with bankruptcy risk are large enough to
be an important component of financial distress costs (Titman, 1984; Berk, Stanton, and Zechner,
2010), which may help to explain the debt conservatism puzzle.

5. Conclusion
We examine how corporate bankruptcy leads to employee earnings losses due to
adjustments in labor markets. We use worker-firm matched data from the U.S. Census Bureau to
show that employee earnings deteriorate significantly post-bankruptcy. The present value of
earnings losses accumulated over seven years averages 87% of pre-bankruptcy annual earnings.
Earnings losses are larger for individuals who leave the industry, and who work for smaller firms
or in thinner local labor markets. These results highlight the role of industry-specific human capital
and costs of moving across geographical areas as important forces behind the post-bankruptcy
earnings loss.
Given that a firm’s debt choice affects the probability of bankruptcy, we argue that these
ex-post lost wages circle back to high-leverage companies in the form of employees demanding
higher ex ante wages (wage premia). We show that the ex-ante wage premium that firms must pay
to compensate for the expected earnings loss due to bankruptcy is a considerable fraction of the
magnitude of the tax benefits of debt. Thus, our results suggest that these indirect bankruptcy costs
are significant enough to be a first-order consideration when firms make capital structure choices.
An important future research direction is to provide direct evidence of the causal impact of wage
premia on actual capital structure choices.

36

References:
Abowd, John M. and Orley C. Ashenfelter, 1981, Anticipated Unemployment, Temporary Layoffs, and
Compensating Wage Differentials, NBER Chapters, in: Studies in Labor Markets, pp. 141–186,
National Bureau of Economic Research.
Abowd, John M., Francis Kramarz, and David N. Margolis, 1999, High Wage Workers and High Wage
Firms, Econometrica 67, 251–333.
Abowd, John M., Bryce E. Stephens, Lars Vilhuber, Fredrik Andersson, Kevin L. McKinney, Marc Roemer,
and Simon Woodcock, 2009, The LEHD Infrastructure Files and the Creation of the Quarterly
Workforce Indicators, in: T. Dunne, J.B. Jensen and M.J. Roberts, eds., Producer Dynamics: New
Evidence from Micro Data (Chicago: University of Chicago Press for the National Bureau of
Economic Research, 2009), pp. 149–230.
Agrawal, Ashwini K. and David A. Matsa, 2013, Labor Unemployment Risk and Corporate Financing
Decisions, Journal of Financial Economics 108, 449–470.
Almeida, Heitor and Thomas Philippon, 2007, The Risk-Adjusted Cost of Financial Distress, Journal of
Finance 62, 2557–2586.
Andrade, Gregor and Steven N. Kaplan, 1998, How Costly is Financial (not Economic) Distress? Evidence
from Highly Levered Transactions that Became Distressed, Journal of Finance 53, 1443–1493.
Autor, David H., David Dorn, Gordon H. Hanson, and Jae Song, 2014, Trade Adjustment: Worker-Level
Evidence, Quarterly Journal of Economics 129, 1799–1860.
Azariadis, Costas, 1975, Implicit Contracts and Underemployment Equilibria, Journal of Political
Economy 83, 1183–1202.
Baghai, Ramin, Rui Silva, Viktor Thell, and Vikrant Vig, 2021, Talent in Distressed Firms: Investigating
the Labor Costs of Financial Distress, Journal of Finance 76, 2907-2961.
Baily, Martin N., 1974, Wages and Employment under Uncertain Demand, Review of Economic Studies 41,
37–50.
Becker, Gary S., 1962, Investment in Human Capital: A Theoretical Analysis, Journal of Political Economy
70, 9–49.
Ben-David, Itzhak, John R. Graham, and Campbell R. Harvey, 2013, Managerial Miscalibration, Quarterly
Journal of Economics 128, 1547–1584.

37

Benmelech, Efraim, Nittai K. Bergman, and Ricardo J. Enriquez, 2012, Negotiating with Labor under
Financial Distress, Review of Corporate Finance Studies 1, 28–67.
Berger, Philip G., Eli Ofek, and David Yermack, 1997, Managerial Entrenchment and Capital Structure
Decisions, Journal of Finance 52, 1411–1438.
Berk, Jonathan, Richard Stanton, and Josef Zechner, 2010, Human Capital, Bankruptcy, and Capital
Structure, Journal of Finance 65, 891–926.
Bernstein, Shai, Emanuele Colonnelli, and Benjamin Iverson, 2019, Asset Allocation in Bankruptcy,
Journal of Finance 74, 5–53.
Blinder, Alan S. and Don H. Choi, 1990, A Shred of Evidence on Theories of Wage Stickiness, Quarterly
Journal of Economics 105, 1003–1015.
Brown, Charles, 1980, Equalizing Differences in the Labor Market, Quarterly Journal of Economics 94,
113–134.
Brown, Jennifer and David Matsa, 2016, Boarding a Sinking Ship? An Investigation of Job Applications to
Distressed Firms, Journal of Finance 71, 507–550.
Butt-Jaggia, Priscilla and Anjan V. Thakor, 1994, Firm-Specific Human Capital and Optimal Capital
Structure, International Economic Review 35, 283–308.
Campbell, Carl M. and Kunal S. Kamlani, 1997, The Reasons for Wage Rigidity: Evidence from a Survey
of Firms, Quarterly Journal of Economics 112, 759–789.
Chava, Sudheer and Michael R. Roberts, 2008, How Does Financing Impact Investment? The Role of Debt
Covenants, Journal of Finance 63, 2085–2121.
Chemmanur, Thomas J., Yingmei Cheng, and Tianming Zhang, 2013, Human Capital, Capital Structure,
and Employee Pay: An Empirical Analysis, Journal of Financial Economics 110, 478–502.
Couch, Kenneth A. and Dana W. Placzek, 2010, Earnings Losses of Displaced Workers Revisited,
American Economic Review 100, 572–589.
Cronqvist, Henrik, Fredrik Heyman, Mattias Nilsson, Helena Svaleryd, and Jonas Vlachos, 2009, Do
Entrenched Managers Pay Their Workers More? Journal of Finance 64, 309–339.
Davis, Steven J. and Till von Wachter, 2011, Recessions and the Costs of Job Loss, Brookings Papers on
Economic Activity, 1–72.
Diamond, Peter A., 1982, Aggregate Demand Management in Search Equilibrium, Journal of Political
Economy 90, 881–894.

38

Dickens, William T., Lorenz Goette, Erica L. Groshen, Steinar Holden, Julian Messina, Mark E. Schweitzer,
Jarkko Turunen, and Melanie E. Ward, 2007, How Wages Change: Micro Evidence from the
International Wage Flexibility Project, Journal of Economic Perspectives 21, 195–214.
Dore, Timothy E. and Rebecca Zarutskie, 2018, Firm Leverage, Labor Market Size, and Employee Pay,
Working paper, Federal Reserve Board.
Eckbo, B. Espen, Karin S. Thorburn, and Wei Wang, 2016, How Costly is Corporate Bankruptcy for the
CEO?, Journal of Financial Economics 121, 210–229.
Elkamhi, Redouane, Jan Ericsson, and Christopher A. Parsons, 2012, The Cost and Timing of Financial
Distress, Journal of Financial Economics 105, 62–81.
Falato, Antonio and J. Nellie Liang, 2016, Do Creditor Rights Increase Employment Risk? Evidence from
Loan Covenants, Journal of Finance 71, 2545–2590.
Gelman, Andrew, and Guido Imbens, 2019, Why High-Order Polynomials Should Not Be Used in
Regression Discontinuity Designs, Journal of Business and Economic Statistics 37, 447–456.
Gilson, Stuart, 1989, Management Turnover and Financial Distress, Journal of Financial Economics 25,
241–262.
Goyal, Vidhan K. and Wei Wang, 2017, Provision of Management Incentives in Bankrupt Firms, Journal
of Law, Finance, and Accounting 2, 87–123.
Graham, John R., 2000, How Big Are the Tax Benefits of Debt?, Journal of Finance 55, 1901–1941.
Graham, John R., 2022, Presidential Address: Corporate Finance and Reality, Working paper, Duke
University.
Guiso, Luigi, Luigi Pistaferri, and Fabiano Schivardi, 2005, Insurance within the Firm, Journal of Political
Economy 113, 1054–1087.
He, Alex, and Daniel le Maire, 2020, Mergers and Managers: Manager-Specific Wage Premiums and Rent
Extraction in M&As, Working paper, University of Maryland and University of Copenhagen.
Hotchkiss, Edith S., 1995, Postbankruptcy Performance and Management Turnover, Journal of Finance 50,
3–21.
Hummels, David, Rasmus Jorgensen, Jakob Munch, and Chong Xiang, 2014, The Wage Effects of
Offshoring: Evidence from Danish Matched Worker-Firm Data, American Economic Review 104,
1597–1629.

39

Imbens, Guido W. and Donald B. Rubin, 2015, Causal Inference for Statistics, Social, and Biomedical
Sciences: An Introduction. New York: Cambridge University Press.
Jacobson, Louis S., Robert J. LaLonde, and Daniel G. Sullivan, 1993, Earnings Losses of Displaced
Workers, American Economic Review 83, 685–709.
Jiang, Wei, Kai Li, and Wei Wang, 2012, Hedge Funds and Chapter 11, Journal of Finance 67, 513–560.
Kambourov, Gueorgui and Iourii Manovskii, 2009, Occupational Specificity of Human Capital,
International Economic Review 50, 63–115.
Kim, Hyunseob, 2020, How Does Labor Market Size Affect Firm Capital Structure? Evidence from Large
Plant Openings, Journal of Financial Economics 138, 277–294.
Kodrzycki, Yolanda K., 2007, Using Unexpected Recalls to Examine the Long-Term Earnings Effects of
Job Displacement, Working paper, Federal Reserve Bank of Boston.
Lazear, Edward P. and Paul Oyer, 2004, Internal and External Labor Markets: A Personnel Economics
Approach, Labour Economics 11, 527–554.
LoPucki, Lynn M. and William C. Whitford, 1992, Patterns in the Bankruptcy Reorganization of Large,
Publicly Held Companies, Cornell Law Review 78, 597–618.
Malmendier, Ulrike, Geoffrey Tate, and Jonathan Yan, 2011, Overconfidence and Early-Life Experiences:
The Effect of Managerial Traits on Corporate Financial Policies, Journal of Finance 66, 1687–1733.
Manning, Alan, 2003, Monopsony in Motion: Imperfect Competition in Labor Markets, Princeton
University Press: Princeton, NJ.
Manning, Alan and Barbara Petrongolo, 2017, How Local Are Labour Markets? Evidence from a Spatial
Job Search Model, American Economic Review 83, 2877–2907.
Matsa, David A., 2018, Capital Structure and a Firm's Workforce, Annual Review of Financial Economics
10, 387–412.
Moretti, Enrico, 2011, Local labor markets, Handbook of Labor Economics, Elsevier, N.H.
Neal, Derek, 1995, Industry-Specific Human Capital: Evidence from Displaced Workers, Journal of Labor
Economics 13, 653–677.
Petrongolo, Barbara and Christopher Pissarides, 2006, Scale Effects in Markets with Search, Economic
Journal 116, 21–44.
Rosen, Sherwin, 1986, The theory of equalizing differences, Handbook of Labor Economics, Elsevier, N.H.

40

Schoeni, Robert F. and Michael Dardia, 2003, Estimates of Earnings Losses of Displaced Workers Using
California Administrative Data, Working paper, RAND and Public Policy Institute of California.
Tate, Geoffrey and Liu Yang, 2015, The Bright Side of Corporate Diversification: Evidence from Internal
Labor Markets, Review of Financial Studies 28, 2203–2249.
Titman, Sheridan, 1984, The Effect of Capital Structure on a Firm’s Liquidation Decision, Journal of
Financial Economics 13, 137–151.
Topel, Robert H., 1984, Equilibrium Earnings, Turnover, and Unemployment: New Evidence, Journal of
Labor Economics 2, 500–522.
Vilhuber, Lars and Kevin L. McKinney, 2014, LEHD Data Documentation LEHD-Overview-S2011:
LEHD Infrastructure Files in the Census RDC – Overview, CES Working paper 14–26.
von Wachter, Till, Jae Song, and Joyce Manchester, 2011, Long-Term Earnings Losses due to Mass Layoffs
During the 1982 Recession: An Analysis Using U.S. Administrative Data from 1974 to 2004,
Working paper, Columbia University, Social Security Administration, and Congressional Budget
Office.
von Wachter, Till, Elizabeth Handwerker, and Andrew K. G. Hildreth, 2009, Estimating the ‘True’ Cost of
Job Loss: Evidence using Matched Data from California 1991-2000, Working paper, Columbia
University, Bureau of Labor Statistics, and University of California at Berkeley.
Walker, W. Reed, 2013, The Transitional Costs of Sectoral Reallocation: Evidence from the Clean Air Act
and the Workforce, Quarterly Journal of Economics 128, 1787–1835.

41

Appendix - Definition of Variables
Variable Name

Variable Definition

Firm characteristics
Sales
Book assets
Book (Market) leverage
ROA

Total sales of the firm in $millions
Total book value of assets in $millions
Total debt + market equity, where total debt = long term plus short-term debt, in
$millions
Total debt/book assets (book assets – book equity + market equity)
Operating income before depreciation and amortizations/lagged book assets

Market-to-book

(Total debt + market equity)/(total debt + book equity)

Number of employees

Number of employees obtained from Longitudinal Business Database (LBD)

Credit rating

S&P credit rating from Compustat

Asset tangibility

Net PP&E/book assets

Capex/Net PP&E

Capital expenditures/lagged net PP&E

R&D/Assets

R&D expense/lagged book assets

Wages/Assets

Wage bills (LBD)/book assets

Worker characteristics

From LEHD-ICF

Female

An indicator variable equal to one if the worker is female, and zero otherwise

Experience

Years of potential work experience = age – (imputed years of education + 6)

Imputed education

Years of education imputed by the Census Bureau

Market value of assets

Dependent and main independent variables
Log (earnings)
BR

Natural logarithm of annual (or quarterly) earnings from LEHD-EHF adjusted for
inflation using the CPI (in 2001 constant dollars)
An indicator variable equal to one for employees of bankrupt firms, and zero for
employees of control firms

d[k], where k = -3 to +6

Event year indicator variables from three years before to six years after bankruptcy

Earnings growth

Change in natural logarithm of annual earnings (relative to the previous year) from
LEHD-EHF adjusted for inflation using the CPI (in 2001 constant dollars)

Bind

An indicator variable equal to one if the worker is an employee of a firm that
violates a loan covenant concerning the current ratio or net worth

Distance (NW or CR)
2nd order distance NW & CR

Relative distance of the firm’s net worth (or current ratio) from the covenant
threshold
Squared relative distances of the firm’s net worth and current ratio

42

Figure 1: Effect of Corporate Bankruptcy on Employee Earnings
Panel A of the figure uses the difference-in-differences estimates in Table 4, Panel A, column (5) and plots the log
change in real earnings without unemployment insurance (UI) adjustments (solid line) for employees of bankrupt
firms from the benchmark (i.e., t-4) earnings, relative to the earnings change for employees of matched firms. Panel
B plots the cumulative change in real earnings without UI adjustment. The dashed lines represent 95% confidence
intervals. “Year t” represents the year of bankruptcy filing.
Panel A: Change in Log Earnings
0.25
0.20

Log earnings change

0.15
0.10
0.05
0.00
t-3

t-2

t-1

t

t+1

t+2

t+3

t+4

t+5

t+6

-0.05
-0.10
-0.15
-0.20
-0.25
-0.30

Years relative to bankruptcy

Panel B: Cumulative Change in Log Earnings

Cumulative log earnings changes

0.35
0.15
-0.05

t-3

t-2

t-1

t

t+1

t+2

t+3

-0.25
-0.45
-0.65
-0.85
-1.05
-1.25
-1.45
-1.65

Years relative to bankruptcy

43

t+4

t+5

t+6

Table 1: Coverage of LEHD States and Years
This table presents the coverage of states and years by the U.S. Census Bureau’s Longitudinal Employer-Household
Dynamics (LEHD)-Employment History File (EHF). See Vilhuber and McKinney (2014) for details of the LEHD
infrastructure.

State
Arkansas
California
Colorado
Florida
Georgia
Hawaii
Iowa
Idaho
Illinois
Indiana
Louisiana
Maryland
Maine
Montana
North Carolina
Nevada
New Jersey
New Mexico
Oklahoma
Oregon
Rhode Island
South Carolina
Texas
Tennessee
Utah
Virginia
Vermont
Washington
Wisconsin
West Virginia

State abbreviation
AR
CA
CO
FL
GA
HI
IA
ID
IL
IN
LA
MD
ME
MT
NC
NV
NJ
NM
OK
OR
RI
SC
TX
TN
UT
VA
VT
WA
WI
WV

44

First year
2002
1991
1990
1992
1994
1995
1998
1990
1990
1990
1990
1985
1996
1993
1991
1998
1996
1995
2000
1991
1995
1998
1995
1998
1999
1998
2000
1990
1990
1997

Last year
2008
2008
2008
2008
2008
2008
2008
2008
2008
2008
2008
2008
2008
2008
2008
2008
2008
2008
2008
2008
2008
2008
2008
2008
2008
2008
2008
2008
2008
2008

Table 2: Descriptive Statistics on Characteristics of Bankrupt and Matched Firms
This table presents descriptive statistics on characteristics of bankruptcy events, firms and their matched firms. Panel A presents bankruptcy outcomes for the full
sample of the bankrupt firms in the LoPucki Bankruptcy Research Database (BRD) from 1990 to 2005 and the matched bankrupt firms with at least ten employees
in the Census Bureau’s Longitudinal Employer-Household Dynamics (LEHD) database. Panel B compares firm characteristics of bankrupt and propensity-score
matched firms, with the statistics based on the values in the latest fiscal year before bankruptcy (in year t-1 or t-2, where “year t” is the year of bankruptcy filing).
Panel C presents the descriptive statistics on dynamics of firm characteristics for bankrupt and matched firms from years t-3 to t-1. See the Appendix for variable
definitions. Statistical significance at the 10%, 5%, and 1% levels is indicated by *, **, and ***. “n.a.” indicates cases in which statistics are not available.

Panel A: Bankruptcy Outcomes

Merged, acquired, or continue
Liquidated, closed, or refile Chapter 11
Unknown
Total

All bankrupt firms from BRD
N. events
% of sample
224
43.9%
106
20.8%
180
35.3%
510
100%

Panel B: Firm Characteristics of Bankrupt and Matched Firms
Bankrupt firms with
Bankrupt firms with
Sample
information from
employees in LEHD
Compustat/LBD
(1)
(2)
(3)
(4)
Variable
Mean
STD
Mean
STD
Sales ($m)
1,537
6,113
1,710
3,693
Book assets ($m)
1,390
4,244
1,278
1,781
Market value of assets ($m)
1,185
4,333
954
1,277
Log market value of assets
6.26
1.05
6.32
1.00
∆ log market value of assets
-0.19
0.48
-0.16
0.41
Number of employees
8,672
20,920
9,497
22,940
Book leverage
0.63
0.37
0.60
0.35
Market leverage
0.57
0.26
0.57
0.23
ROA
0.00
0.31
0.05
0.09
Market-to-book
1.18
0.87
1.09
0.66
Wages/Assets
0.22
0.25
0.25
0.24
Capex/Net PP&E
0.32
0.47
0.25
0.32
R&D/Assets
0.01
0.03
0.01
0.02
N. events
360
130
-

45

Matched bankrupt firms with employees in LEHD
N. events
% of sample
40
30.8%
30
23.1%
60
46.2%
130
100%

Propensity-score
matched firms
(5)
Mean
1,460
1,224
1,168
6.08
-0.14
6,067
0.35
0.32
0.06
1.12
0.22
0.24
0.01
130

(6)
STD
3,384
2,381
3,814
1.16
0.41
10,270
0.28
0.23
0.08
0.46
0.15
0.26
0.03
-

t-statistic for
difference
(7)
(3)–(1)
0.51
-0.46
-0.95
0.89
0.70
0.53
-1.13
-0.30
3.05***
-1.75*
1.62
-2.49**
-0.60
-

(8)
(3)–(5)
0.56
0.19
-0.58
1.63
-0.44
1.52
6.57***
8.38***
-0.82
-0.47
1.20
0.18
-0.92
-

Panel C: Dynamics of Firm Characteristics for Bankrupt and Matched Firms Prior to Bankruptcy
(2)
(1)
Bankrupt firms (N = 90)
Matched firms (N = 90)
Year
t-3
t-2
t-1
t-3
t-2
t-1
Sales ($m)
1762
1871
1799
1,375
1,524
1,512
Book assets ($m)
1437
1497
1380
1,246
1,389
1,196
Market value of assets ($m)
1402
1175
952
1,623
1,232
902
Log market value of assets ($m)
6.61
6.54
6.36
6.31
6.23
6.10
∆ log market value of assets ($m,
from t-3)
n.a.
-0.08
-0.26
n.a.
-0.08
-0.22
Number of employees
12,860
12,430
10,980
7,771
7,270
6,741
Book leverage
0.43
0.51
0.62
0.31
0.33
0.36
Market leverage
0.37
0.46
0.58
0.26
0.30
0.33
ROA
0.11
0.09
0.05
0.12
0.09
0.06
Market-to-book
1.57
1.25
1.04
1.49
1.19
1.09
Wages/Assets
0.27
0.26
0.26
0.23
0.22
0.22
Capex/Net PP&E
0.35
0.28
0.18
0.40
0.31
0.21
R&D/Assets
0.01
0.01
0.01
0.01
0.01
0.01

46

t-statistic for (1)–(2)
t-3

t-2

t-1

0.80
0.56
-0.30
1.59

0.65
0.26
-0.13
1.70*

0.52
0.58
0.21
1.58

n.a.
1.46
3.22***
3.12***
-0.40
0.49
1.41
-0.63
-0.41

0.07
1.60
4.73***
5.00***
-0.40
0.71
1.32
-0.59
-0.30

-0.48
1.40
5.86***
7.10***
-1.02
-0.94
1.19
-1.10
0.13

Table 3: Descriptive Statistics on Characteristics of Workers Employed by Bankrupt and Control Firms
This table presents descriptive statistics of the workers employed by bankrupt and control firms. All variables are measured at t-1 (where “year t” is the year of
bankruptcy filing). The earnings data for individual employees are from the LEHD-EHF (Employment History Files), and other individual characteristics are from
the LEHD-ICF (Individual Characteristics Files). We require that workers have at least two years of tenure and ages between 22 and 55 years in year t-1. Columns
(5) and (6) show statistics for a 0.5% random sample of workers from the LEHD-EHF who are employed by non-bankrupt public firms, and satisfy the same
requirements for industry (i.e., excluding financials and utilities), tenure, and age as the workers in the bankruptcy sample. Earnings are CPI-adjusted (in 2001
constant dollars). “% stay in firm (before)” (industry, county, or LEHD) is the percent of employees who stay in the firm (the employee’s industry, county, or
LEHD) from t-4 to t-1. Similarly, “% stay in firm (after)” (industry, county, or LEHD) is the percent of employees who stay in the firm (the employee’s industry,
county, or LEHD) from t-1 to t+3. “Δ % stay in firm (before → after)” (industry, county, or LEHD) is the change in the percent of employees who stay in the firm
(the employee’s industry, county, or LEHD) between t-4 to t-1 and t-1 to t+3. In columns (5) and (6), we report “% stay in firm/industry/county (before)” from t-1
to t-4 (instead of from t-4 to t-1) because the randomly selected worker sample does not condition on firms but on individuals. As a result, a sample of workers
who were employees of the “same firm/industry/county” several years before t-1 is not well defined. For a similar reason, we are not able to report “% stay in
LEHD (before).” Definitions of variables are in the Appendix. Columns (7) and (8) report heteroskedasticity robust t-statistics adjusted for within-firm clustering.
Statistical significance at the 10%, 5%, and 1% levels is indicated by *, **, and ***. The number of employees is rounded to the nearest thousand to follow the Census
Bureau’s disclosure rules. “n.a.” indicates cases in which statistics are not available.

Years of education (imputed)
Age
Years of experience
Female
Annual real earnings at t-1 (2001$)
Number of employees
% stay in firm (before)
% stay in industry (before)
% stay in county (before)
% stay in LEHD (before)
% stay in firm (after)
% stay in industry (after)
% stay in county (after)

Employees of
bankrupt firms
(1)
(2)
Mean
STD
13.34
2.44
37.64
9.35
18.30
9.21
0.51
0.50
31,460
33,430
234,000
54.6
49.8
63.7
48.1
66.8
47.1
66.9
47.1
32.4
49.5
52.9
48.8
53.4
48.1

Employees of
matched firms
(3)
(4)
Mean
STD
13.29
2.36
36.88
9.50
17.59
9.34
0.46
0.50
28,710
94,210
154,000
51.4
50.0
63.1
48.3
66.6
47.2
63.9
48.0
43.2
46.8
61.1
49.9
63.7
49.9

47

Randomly
selected workers
(5)
(6)
Mean
STD
13.46
2.41
37.17
9.37
17.71
9.21
0.45
0.50
35,100
83,960
267,000
68.5
76.9
77.2
n.a.
58.0
70.2
70.9

46.5
42.2
41.9
n.a.
49.4
45.7
45.4

t-statistic
for (1)–(3)

t-statistic
for (1)–(5)

(7)

(8)

0.42
1.40
1.39
0.84
0.55
0.51
0.12
0.03
0.53
-1.66*
-2.10**
-2.61***

-1.04
1.03
1.45
0.93
-0.75
-3.03***
-2.86***
-2.60**
n.a.
-4.75***
-5.69***
-5.73***

% stay in LEHD (after)
Δ % stay in firm (before → after)
Δ % stay in industry (before → after)
Δ % stay in county (before → after)
Δ % stay in LEHD (before → after)

57.3
-22.1
-10.8
-13.4
-9.6

48.6
n.a.
n.a.
n.a.
n.a.

61.6
-8.2
-2.0
-2.9
-2.4

48

49.5
n.a.
n.a.
n.a.
n.a.

66.3
-10.5
-6.7
-6.3
n.a.

47.3
n.a.
n.a.
n.a.
n.a.

-1.56
-1.82*
-1.86*
-2.57**
-1.20

-3.80***
-2.02**
-1.17
-2.91***
n.a.

Table 4: Effect of Corporate Bankruptcy on Employee Earnings
This table presents difference-in-differences estimates of earnings changes for workers employed by bankrupt firms
around bankruptcy filings relative to a control group of workers. The control group in Panel A includes employees of
the matched firms, and in Panel B a 0.5% random sample of workers from the LEHD universe employed by public
non-bankrupt firms. Panel C reports the present value of earnings losses after estimated unemployment insurance (UI)
income is incorporated. The dependent variable is log(annual earnings) in 2001 constant dollars. “BR” is an indicator
variable equal to one for employees of bankrupt firms. The event year indicator variables are “d[k],” where -3 ≤ k ≤
6. The regressions use the observations from event years t-4 to t+6, with the benchmark earnings being the earnings
in year t-4. Standalone event year indicators d[k] (-3 ≤ k ≤ 6) are included in all regressions but suppressed for
expositional convenience. “PV (Earnings losses)” (“PV (Earnings losses, Rf)”) is the present value of earnings losses
from years t to t+6 computed using a real risk-adjusted discount rate (real risk-free rate) as a percent of the prebankruptcy annual earnings in t-4. To illustrate the process to compute the PV, take Panel A, column (5) as an example.
We first convert the coefficient on “d[1] × BR,” -0.153, into exp(-0.153) – 1 = -14.2%, which is the percentage change
in earnings when “d[1] × BR” changes from zero to one. We then discount the percentage change by a year using a
real discount rate of -7.40% or the risk-free rate of 3.34%. The real discount rate is inflation- and risk-adjusted that is
appropriate given workers’ exposure to uninsurable idiosyncratic risk. The risk-free rate ignores the worker’s
(idiosyncratic) risk exposure. We repeat the calculation for all coefficients on “d[k] × BR” for 0 ≤ k ≤ 6, and then sum
up the present values across the seven years. For regression coefficients, heteroskedasticity robust t-statistics adjusted
for within-firm and worker clustering are in parentheses. Definitions of variables are in the Appendix. For present
values of earnings losses, t-statistics for the PV significantly differing from zero are in parentheses (standard errors
are computed using the delta method). Statistical significance at the 10%, 5%, and 1% levels is indicated by *, **, and
***
. The number of observations is rounded to the nearest thousand to follow the Census Bureau’s disclosure rules.

49

Panel A: Earnings Losses Using Matched Firms’ Employees as Control Group
Dep. Var. = Log (Earnings)
(1)
(2)
(3)
(4)
-0.004
0.033
0.024
0.001
d[-3] × BR
(-0.17)
(1.31)
(1.29)
(0.05)
d[-2] × BR
-0.034
0.043
-0.004
-0.017
(-0.83)
(1.00)
(-0.11)
(-0.39)
-0.034
0.038
0.006
0.007
d[-1] × BR
(-0.81)
(0.83)
(0.16)
(0.14)
d[0] × BR
-0.086**
-0.045
-0.083***
-0.106***
(-2.43)
(-1.36)
(-2.93)
(-3.78)
-0.159***
-0.105***
-0.126***
-0.161***
d[1] × BR
(-3.39)
(-2.66)
(-3.69)
(-5.15)
d[2] × BR
-0.126**
-0.144***
-0.055
-0.104***
(-2.20)
(-3.26)
(-1.63)
(-3.03)
-0.037
-0.079**
-0.013
-0.097***
d[3] × BR
(-0.51)
(-2.08)
(-0.30)
(-2.61)
d[4] × BR
-0.056
-0.044
-0.030
-0.074*
(-0.87)
(-1.10)
(-0.75)
(-1.76)
-0.014
0.026
0.017
-0.031
d[5] × BR
(-0.18)
(0.71)
(0.32)
(-0.89)
d[6] × BR
-0.046
0.012
0.017
-0.033
(-0.61)
(0.28)
(0.43)
(-0.76)
Female × Experience
0.010***
-0.001
0.005**
-0.002*
(2.83)
(-1.02)
(2.65)
(-1.68)
Education × Experience
-0.005***
-0.004***
-0.004***
-0.004***
(-11.69)
(-19.70)
(-19.22)
(-22.19)
Standalone event year
Yes
Yes
Yes
Yes
indicators d[k] (-3 ≤ k ≤ 6)
Worker FE
Yes
Yes
Yes
Yes
Year FE
Yes
SIC2 × Year FE
Yes
Yes
County × Year FE
Yes
Yes
SIC2 × County × Year FE
Observations
2,967,000
2,967,000
2,967,000
2,967,000
R2
57.1%
58.3%
59.3%
59.9%
PV (Earnings losses)
-58.9%
-40.2%
-27.7%
-68.7%**
(-1.24)
(-1.41)
(-0.89)
(-2.51)
PV (Earnings losses, Rf)
-46.4%
-33.9%*
-25.6%
-53.6%***
(-1.45)
(-1.71)
(-1.21)
(-2.88)

50

(5)
0.018
(0.56)
-0.012
(-0.19)
0.001
(0.01)
-0.128***
(-3.55)
-0.153***
(-3.89)
-0.104**
(-2.29)
-0.101
(-1.63)
-0.111
(-1.45)
-0.067
(-1.16)
-0.098
(-1.46)
-0.002
(-1.56)
-0.004***
(-22.41)
Yes
Yes

Yes
2,967,000
61.4%
-89.5%**
(-2.14)
-66.0%**
(-2.35)

Panel B: Earnings Losses Using Randomly Selected Workers as Control Group
Dep. Var. = Log (Earnings)
(1)
(2)
(3)
(4)
d[-3] × BR
0.008
-0.001
0.007
-0.015
(0.42)
(-0.08)
(0.64)
(-1.17)
d[-2] × BR
0.004
-0.005
0.017
-0.017
(0.12)
(-0.18)
(0.55)
(-0.58)
d[-1] × BR
-0.036
-0.029
-0.001
-0.035
(-1.05)
(-0.91)
(-0.04)
(-1.13)
d[0] × BR
-0.125***
-0.138***
-0.120***
-0.182***
(-3.70)
(-6.00)
(-4.62)
(-9.59)
d[1] × BR
-0.184***
-0.204***
-0.157***
-0.240***
(-4.00)
(-6.80)
(-5.16)
(-9.93)
d[2] × BR
-0.200***
-0.241***
-0.122***
-0.199***
(-4.70)
(-8.16)
(-5.17)
(-8.45)
d[3] × BR
-0.120***
-0.169***
-0.050**
-0.144***
(-2.74)
(-6.85)
(-1.97)
(-5.97)
d[4] × BR
-0.131***
-0.167***
-0.039
-0.133***
(-2.80)
(-5.59)
(-1.44)
(-4.63)
d[5] × BR
-0.050
-0.052
0.019
-0.050
(-1.28)
(-1.26)
(0.65)
(-1.15)
d[6] × BR
-0.053
-0.102***
0.042
-0.067**
(-0.77)
(-2.59)
(1.37)
(-2.12)
Female × Experience
0.011***
-0.003**
0.007***
-0.004***
(4.85)
(-2.50)
(5.25)
(-2.98)
Education × Experience
-0.005***
-0.004***
-0.004***
-0.004***
(-14.71)
(-19.22)
(-22.56)
(-20.86)
Standalone event year
Yes
Yes
Yes
Yes
indicators d[k] (-3 ≤ k ≤ 6)
Worker FE
Yes
Yes
Yes
Yes
Year FE
Yes
SIC2 × Year FE
Yes
Yes
County × Year FE
Yes
Yes
SIC2 × County × Year FE
Observations
3,855,000
3,855,000
3,855,000
3,855,000
R2
59.3%
60.0%
60.9%
61.2%
PV (Earnings losses)
-97.3%***
-121.3%***
-42.2%**
-112.1%***
(-2.81)
(-5.88)
(-2.07)
(-5.68)
PV (Earnings losses, Rf)
-74.4%***
-90.7%***
-38.8%***
-86.8%***
(-3.16)
(-6.6)
(-2.77)
(-6.62)

51

(5)
0.006
(0.32)
0.003
(0.08)
-0.027
(-0.67)
-0.187***
(-8.08)
-0.210***
(-7.80)
-0.180***
(-5.74)
-0.094***
(-2.99)
-0.076**
(-2.04)
0.022
(0.38)
-0.027
(-0.66)
-0.004***
(-3.08)
-0.004***
(-20.66)
Yes
Yes

Yes
3,855,000
63.9%
-78.9%***
(-2.96)
-65.7%***
(-3.71)

Panel C: Earnings Losses Adjusted for Unemployment Insurance Income
Dep. Var. = Log (Earnings)
(1)
(2)
(3)
(4)
(5)
Worker-level controls
Yes
Yes
Yes
Yes
Yes
Worker FE
Yes
Yes
Yes
Yes
Yes
Year FE
Yes
SIC2 × Year FE
Yes
Yes
County × Year FE
Yes
Yes
SIC2 × County × Year FE
Yes
Using Matched Firms’ Employees as Control Group
PV (Earnings losses)
-54.5%
-33.9%
-26.3%
-64.4%**
-87.1%**
(-1.15)
(-1.17)
(-0.84)
(-2.31)
(-2.07)
PV (Earnings losses, Rf)
-42.9%
-28.7%
-24.2%
-50.2%**
-63.8%**
(-1.35)
(-1.43)
(-1.14)
(-2.62)
(-2.23)
Using Randomly Selected Workers as Control Group
PV (Earnings losses)
-94.6%*** -118.0%***
-40.2%**
-110.0%***
-80.2%***
(-2.82)
(-5.81)
(-1.99)
(-5.65)
(-3.02)
PV (Earnings losses, Rf)
-71.6%***
-87.4%***
-36.5%**
-84.2%***
-65.1%***
(-3.15)
(-6.39)
(-2.66)
(-6.44)
(-3.66)

52

Table 5: Effect of Corporate Bankruptcy on Employee Earnings - Conditional Analysis on
Worker Mobility, Labor Market Characteristics, and Bankruptcy Liquidation Outcomes
This table presents the present value of earnings losses conditional on worker mobility, labor market characteristics,
and bankruptcy liquidation outcomes. “PV (Earnings losses)” is the present value of earnings losses from years t to
t+6 (computed using a real risk-adjusted discount rate of -7.40%) as a percent of the pre-bankruptcy annual earnings
in t-4. Panel A reports the present values of earnings losses conditional on worker mobility. We define a worker as a
stayer if she stays in the same firm, industry, or county through year t+3 as her t-1 firm, industry, or county. This panel
stratifies the treated (i.e., bankrupt) firm employees based on whether they stay with their firms, industries, or counties,
but not the matched control firm employees. Thus, the panel decomposes earnings losses of the treated group workers
relative to the average worker in the control group. Panel B reports the present values of earnings losses conditional
on firm size (as measured by book assets from Compustat), local labor market size (total employment in a given twodigit SIC industry and county cell), and bankruptcy outcomes (liquidation vs. non-liquidation). We split the sample at
the 75th percentile of the firm size and labor market size distributions. This panel stratifies both the treated and control
firm employees. t-statistics for the (difference in) present value significantly differing from zero are in parentheses
(standard errors are computed using the delta method). Statistical significance at the 10%, 5%, and 1% levels is
indicated by *, **, and ***.

Panel A: Present Value of Earnings Losses Conditional on Worker Mobility
(1)
(2)
(3)
(4)
(5)
Leavers:
Leavers:
Leavers:
Firm
Group:
Firm leavers same industry diff. industry same industry
stayers
same county
same county
diff. county
37.9%
-119.9%***
79.1%
19.0%
8.5%
PV (Earnings losses)
(0.62)
(-3.11)
(1.42)
(0.39)
(0.16)
Diff. between columns
(2)–(1)***
(6)–(4)***
(4)–(3)**
(5)–(3)***
(t-statistic)
(-3.62)
(-8.83)
(-2.45)
(-3.25)
Controls
Yes
Yes
Yes
Yes
Yes
Worker FE
Yes
Yes
Yes
Yes
Yes
SIC2 × County × Year
Yes
Yes
Yes
Yes
Yes
FE

(6)
Leavers:
diff. industry
diff. county
-211.1%***
(-6.02)
(6)–(5)***
(-12.29)
Yes
Yes
Yes

Panel B: Present Value of Earnings Losses Conditional on Firm Size, Labor Market Size, and
Bankruptcy Outcomes
(1)
(2)
(3)
(4)
(5)
(6)
Conditioning variable:
Firm size
Labor market size
Bankruptcy liquidation outcome
Sample:
Large
Small
Large
Small
Liquidation
Non-liquidation
PV (Earnings losses)
-61.5%
-99.1%**
-24.2%
-116.0%**
-97.4%*
-83.6%**
(-1.00)
(-2.32)
(-0.55)
(-2.46)
(-1.75)
(-1.99)
Diff. between two groups
37.6%
91.8%**
-13.8%
(t-statistic)
(0.66)
(2.11)
(-0.33)
Controls
Yes
Yes
Yes
Yes
Yes
Yes
Worker FE
Yes
Yes
Yes
Yes
Yes
Yes
SIC2 × County × Year FE
Yes
Yes
Yes
Yes
Yes
Yes

53

Table 6: Covenant Violations and Employee Earnings Changes - An RDD Analysis
This table presents the effect of loan covenant violations on employee earnings growth using a regression discontinuity
design based on DealScan-Compustat-LEHD matched data from 1995 to 2008. The dependent variable is the change
in a worker’s log(annual earnings) in 2001 constant dollars from the previous year. “Bind” is an indicator variable
equal to one if the worker is an employee of a firm that violates a loan covenant concerning the current ratio or net
worth, and zero otherwise. Worker-level controls include imputed education, female indicator, imputed education ×
female indicator, years of experience, imputed education × years of experience, and female indicator × years of
experience. Distance (NW) is the relative distance of the firm’s net worth from the covenant threshold. Distance (CR)
is the relative distance of the firm’s current ratio from the covenant threshold. “2nd order distance (NW & CR)” is the
squared relative distances of net worth and current ratio. Definitions of variables are in the Appendix.
Heteroskedasticity robust t-statistics adjusted for within-firm and worker clustering are in parentheses. Statistical
significance at the 10%, 5%, and 1% levels is indicated by *, **, and ***. The number of observations is rounded to the
nearest thousands to follow the Census Bureau’s disclosure rules.

Dep. Var. = Earnings growth
Bind
Distance (NW)
Distance (CR)
2nd order distance (NW & CR)
Worker-level controls
SIC2 × county × year FE
Firm FE
Observations
R2

(1)
-0.037**
(-2.13)
0.050***
(8.05)
0.003
(0.31)

(2)
-0.037**
(-2.33)
0.045
(1.47)
0.028
(0.78)

Yes
Yes

Yes
Yes
Yes

1,673,000
8.3%

1,673,000
8.3%

54

(3)
-0.052**
(-2.15)
-0.015
(-0.91)
0.026*
(1.77)

(4)
-0.056**
(-2.21)
-0.020
(-1.35)
-0.001
(-0.04)

Yes
Yes
Yes
1,673,000
8.8%

Yes
Yes
Yes
Yes
1,673,000
8.8%

Table 7: Leverage and Employee Wages - A Wage Regression Approach
This table presents the relationship between corporate leverage and employee earnings. The dependent variable is
log(quarterly earnings) in 2001 constant dollars. The control variables include log(book assets), market-to-book, ROA,
asset tangibility of firms, female indicator × years of experience, imputed education × years of experience of individual
workers, firm, worker, and two-digit SIC industry × county × year fixed effects. To select the sample for the regression,
we begin with Compustat firms with S&P credit ratings between AA+ and B- or unrated, and book assets greater than
$100 million in 1980 constant dollars and less than or equal to the assets of the largest firm in our bankrupt and
matched firm sample from 1986 to 2008. Then we select all workers in LEHD with ages between 22 and 55 and at
least two years of tenure who move between these firms. For earnings before moving, we use the highest quarterly
earnings in the last four quarters at the old firm; for earnings after moving, we use the highest quarterly earnings in
the first four quarters at the new firm. Column (1) presents estimates for the full sample. Columns (2) and (3) present
estimates conditional on firm size, measured by book assets of a given firm. Columns (4) and (5) present estimates
conditional on local labor market size, measured by total employment in a given two-digit SIC industry and county
cell. “Large” (“Small”) represents a subsample in which firm assets or local labor market size is greater than (less than
or equal to) the 75th percentile in the sample distribution. Definitions of variables are in the Appendix.
Heteroskedasticity robust t-statistics adjusted for within-firm and worker clustering are in parentheses. Statistical
significance at the 10%, 5%, and 1% levels is indicated by *, **, and ***. The number of observations is rounded to the
nearest thousands to follow the Census Bureau’s disclosure rules.

Dep. Var. = Log (earnings)
Conditioning variable:
Sample:
Market leverage
Worker-level controls
Firm-level controls
Worker FE
Firm FE
SIC2 × County × Year FE
Observations
R2
Market leverage × (Small – Large)
(t-statistic)

(1)
Full
0.057***
(3.17)
Yes
Yes
Yes
Yes
Yes
9,094,000
70.7%
-

(2)

(3)

Firm size
Small
Large
-0.110
0.066***
(-1.25)
(2.94)
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
2,282,000
6,812,000
91.5%
76.7%
0.176*
(1.94)

55

(4)
(5)
Labor market size
Small
Large
0.014
0.070***
(0.53)
(3.82)
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
2,276,000
6,819,000
84.6%
76.0%
0.056**
(2.05)

Table 8: Implied Wage Premia from Binomial Valuation Model
This table develops implied wage premia for firms with different credit ratings. We use the present value of earnings
losses with unemployment insurance (UI) adjustments as a fraction of pre-bankruptcy annual wages in a binomial
valuation model (see Section A.1 of Online Appendix) to derive the implied wage premia. We first calculate a oneyear risk-adjusted bankruptcy probability q in column (2) as 1-(1-0.6×q 10)1/10, where 0.6 is the estimated probability
of bankruptcy conditional on default (based on Moody’s Default and Recovery Database and the AltmanKuehne/NYU Salomon Center Bond Master Default Database), and q 10 (in column (1)) is the ten-year risk-adjusted
default probability provided in Almeida and Philippon (2007, Table III). We then use a certainty equivalence approach
(described in Online Appendix B) to estimate the one-year risk-adjusted bankruptcy probability for under-diversified
workers, qw, in column (3). In column (4), we use this annual qw to calculate the discounted wage premium a firm
would pay workers as [qw × (NL/W)/(1+rf)], where “NL” represents the present value of earnings losses given
bankruptcy with UI adjustments, net of compensating differentials lost post-bankruptcy and “W” represents prebankruptcy annual wages. The nominal risk-free rate (rf) is assumed to be 6.33%, the average 10-year Treasury rate
from 1985-2008, our sample period. All numbers in the table are in %.

Credit rating

AAA
AA
A
BBB
BB
B
BBB minus AA

q10 = Ten-year riskadjusted default
probability from
Almeida and
Philippon (2007)
(1)
1.65
6.75
12.72
20.88
39.16
62.48
14.13

q = one-year
risk-adjusted
bankruptcy
probability
based on q10
(2)
0.10
0.41
0.79
1.33
2.64
4.59
0.92

56

qw = One-year
risk-adjusted
bankruptcy
probability
(workers)
(3)
0.11
0.46
0.89
1.51
3.02
5.24
1.05

Implied wage premium
as fraction of wages
(Binomial valuation
model)
(4)
0.07
0.29
0.55
0.94
1.89
3.27
0.66

Table 9: Comparison of Wage Premia - Binomial Valuation Model versus Wage Regression
This table compares the implied premia derived from the binomial valuation model in Table 8 with those estimated
from the wage regression in Table 7. In particular, it presents the additional wage premium due to moving from AA
to BBB ratings. Column (1) presents the results for the full sample, and columns (2) through (7) for subsamples by
firm size and labor market size. Firm size is measured by book assets of a given firm, and local labor market size by
total employment in a given two-digit SIC industry and county cell. “Large” (“Small”) represents a subsample in
which firm assets or local labor market size is greater than (less than or equal to) the 75 th percentile in the sample
distribution. Panel A (B) presents the change in implied wage premia (in percent) estimated from the binomial
valuation model (wage regressions in Table 7). Statistical significance, based on estimates from Tables 4 and Online
Appendix A8 (for the valuation model) and from Table 7 (for wage regressions), at the 10%, 5%, and 1% levels is
indicated by *, **, and ***. Heteroskedasticity robust standard errors adjusted for within-firm and worker clustering are
in parentheses.
Firm size

Conditioning variable:
Sample:

Labor market size

Full

Large

Small

Diff.

Large

Small

Diff.

(1)

(2)

(3)

(4)

(5)

(6)

(7)

Panel A: Binomial Valuation Model
PV (wage premia)
BBB minus AA

0.66%**

0.35%

0.74%**

0.39%

0.16%

0.86%**

0.70%**

(Std. Err.)

(0.32%)

(0.46%)

(0.32%)

(0.40%)

(0.33%)

(0.36%)

(0.32%)

Panel B: Wage Regression
PV (wage premia)
BBB minus AA

0.64%***

-1.24%

0.74%***

1.98%*

0.15%

0.79%***

0.64%**

(Std. Err.)

(0.20%)

(0.99%)

(0.25%)

(1.02%)

(0.29%)

(0.21%)

(0.31%)

57

Table 10: Capitalized Value of Implied Wage Premia, Tax Benefits of Financial Leverage, and Expected Costs of Financial
Distress
This table develops the capitalized value of implied wage premia as a fraction of firm value for firms with different credit ratings, and compares them with the tax
benefits of debt and expected costs of financial distress. Panel A assumes that all employees are fully compensated for bankruptcy risk. Column (1) presents the
annual wage premium as a fraction of firm value. Column (2) calculates the present value of implied wage premia as a fraction of firm value that the firm would
pay workers in an infinite horizon (i.e., capitalized value) as [qw/(q+rf)] × NL/MV, where “NL” represents the present value of net earnings losses with
unemployment insurance adjustments given bankruptcy and “MV” represents firm value. The nominal risk-free rate (rf) is assumed to be 6.33%, the average 10year Treasury rate from 1985-2008, our sample period. Online Appendix A provides detailed models underlying these wage premia calculations. Columns (3)-(5)
show the tax benefits of debt and expected costs of financial distress estimated from Almeida and Philippon (2007, Table VI) and Elkamhi, Ericsson, and Parsons
(2012, Tables 2 and 3). Wage premia, tax benefits, and expected costs of financial distress in the table are present values as percentages of pre-bankruptcy firm
value. Panel B presents alternative capitalized values of implied wage premia as a fraction of firm value, assuming that only newly hired employees are compensated
for bankruptcy risk.
Panel A: Assuming All Employees Are Fully Compensated for Bankruptcy Risk

Credit rating

AAA
AA
A
BBB
BB
B
BBB minus AA

Implied annual wage
premium as fraction
of firm value

Capitalized value of
implied wage premium as
fraction of firm value

(1)
0.02
0.07
0.13
0.22
0.44
0.76
0.15

(2)
0.27
1.05
1.91
3.04
5.18
7.38
1.99

Tax benefits of
debt from
Almeida and
Philippon (2007)
(3)
0.47
2.51
4.40
5.18
7.22
8.95
2.67

E(costs of
financial distress)
from Almeida and
Philippon (2007)
(4)
0.32
1.84
3.84
4.53
6.81
9.54
2.69

Panel B: Capitalized Wage Premium Assuming Newly Hired Employees Are Compensated for Bankruptcy Risk

Capitalized value of
implied wage premium as
fraction of firm value
0.50
0.88
1.17

BBB minus AA
New hires in t
New hires in t and t-1
New hires in t, t-1, and t-2

58

E(costs of financial
distress) from
Elkamhi, Ericsson,
Parsons (2012)
(5)
0.0
0.0
0.2
0.5
0.9
2.1
0.5

Employee Costs of Corporate Bankruptcy
Online Appendix (Not for Publication)*
John R. Graham
Hyunseob Kim
Si Li
Jiaping Qiu†
March 2022
Abstract
This online appendix presents a detailed model to compute wage premia (Appendix A), approaches
to account for workers’ exposure to idiosyncratic risk (Appendices B-C), the approach to adjust
earnings losses for unemployment insurance benefits (Appendix D), and supplementary tables and
figures.

*

Any views expressed are those of the authors and not those of the U.S. Census Bureau. The Census Bureau’s
Disclosure Review Board and Disclosure Avoidance Officers have reviewed this information product for unauthorized
disclosure of confidential information and have approved the disclosure avoidance practices applied to this release.
This research was performed at a Federal Statistical Research Data Center under FSRDC Project Number 808
(CBDRB-FY21-P808-R8890, FY22-P808-R9448). This research uses data from the Census Bureau’s Longitudinal
Employer Household Dynamics Program, which was partially supported by the following National Science
Foundation Grants SES-9978093, SES-0339191 and ITR-0427889; National Institute on Aging Grant AG018854;
and grants from the Alfred P. Sloan Foundation.
†
John R. Graham (john.graham@duke.edu), Fuqua School of Business, Duke University, Durham NC 27708, phone:
(919) 660-7857; Hyunseob Kim (hyunseob.kim@chi.frb.org), Economic Research Department, Federal Reserve Bank
of Chicago, Chicago, IL 60604, phone: (312) 322-6131; Si Li (sli@wlu.ca), Financial Services Research Centre,
Lazaridis School of Business and Economics, Wilfrid Laurier University, Waterloo, Ontario N2L 3C5, Canada, phone:
(519) 884-0710 ext. 2395; Jiaping Qiu (qiu@mcmaster.ca), DeGroote School of Business, McMaster University,
Hamilton, Ontario L8S 4M4, Canada, phone: (905) 525-9140 ext. 23963.

1

Online Appendix A
Binomial Valuation Model to Compute Implied Wage Premia Due to Bankruptcy Risk
In a competitive labor market, when an employee expects earnings reductions following a
bankruptcy filing of her employer, she will require higher wages ex ante to compensate for such a
potential loss. In this appendix, we first derive the wage premium demanded by an employee in
Section A.1 and then the capitalized value of wage premia paid by a firm with an infinite life
(going concern) in Section A.2.
A.1. Wage Premium a Worker Demands

p

L

1–p

0

PV(expected earnings loss)

Figure A.1
Let L be the employee’s earnings loss given bankruptcy of her employer, p the historical
annual probability of bankruptcy, and 𝑟 the appropriate discount rate for the worker. Note that L
could be normalized by wage or firm value as in the paper. Then, the present value (PV) of
expected earnings loss in a given year is equal to:
𝑃𝑉(𝑒𝑥𝑝𝑒𝑐𝑡𝑒𝑑 𝑒𝑎𝑟𝑛𝑖𝑛𝑔𝑠 𝑙𝑜𝑠𝑠) =

.

(A-1)

Assuming that the employee is risk averse and the firm is more likely to file for bankruptcy
in bad times, the discount rate 𝑟 < 𝑟 , the risk-free rate. Because we do not observe 𝑟 , we adopt
a risk-neutral approach proposed in Almeida and Philippon (2007) to estimate the PV expected
earnings loss. Specifically,

2

𝑃𝑉(𝑒𝑥𝑝𝑒𝑐𝑡𝑒𝑑 𝑒𝑎𝑟𝑛𝑖𝑛𝑔𝑠 𝑙𝑜𝑠𝑠) =

(A-2)

,

where 𝑞 is the annual risk-adjusted probability of bankruptcy for the worker and rf is the riskfree rate.
If we use a risk-free firm (i.e., 𝑞 = 0) as the benchmark case, in a competitive labor market
the risky firm has to offer the same expected annual wages, W1 – PV(expected earnings loss), to
the employee, the same as the wages offered by a risk-free firm W0, all else equal. This implies
that the annual wage premium demanded by the employee relative to a risk-free firm is:
𝑎𝑛𝑛𝑢𝑎𝑙 𝑤𝑎𝑔𝑒 𝑝𝑟𝑒𝑚𝑖𝑢𝑚 = 𝑊 − 𝑊
= 𝑃𝑉(𝑒𝑥𝑝𝑒𝑐𝑡𝑒𝑑 𝑒𝑎𝑟𝑛𝑖𝑛𝑔𝑠 𝑙𝑜𝑠𝑠) =

.

(A-3)

This result is intuitive; the annual wage premium is equal to the marginal increase in the PV
expected earnings loss resulting from a marginal increase in annual bankruptcy probability.

A.2. Capitalized Value of Wage Premia the Firm Pays a Worker

q

L

L

q
1-q
0

q

L

1-q
0
1-q

0…
Figure A.2
3

In an infinite-period model, the firm will have an employee indefinitely. If we use a risk-free
firm as the benchmark, in year 0, the firm faces a risk-adjusted bankruptcy probability 𝑞 and will
need to offer an annual wage premium of

over a risk-free firm as shown in equation (A-3).3

Note that the risk-neutral probability of bankruptcy for the firm q and the worker 𝑞 are different
as they have different risk preferences (e.g., portfolios). If the firm survives to the next year (with
probability 1 – q), the employee again will face a risk-adjusted bankruptcy probability 𝑞 in that
year, and the firm will again pay an annual wage premium of

over a risk-free firm.

Continuing this process indefinitely, the PV of wage premia that the firm has to offer over a
risk-free firm over its life is:
𝑃𝑉 𝑜𝑓 𝑤𝑎𝑔𝑒 𝑝𝑟𝑒𝑚𝑖𝑎 = ∑

𝑎𝑛𝑛𝑢𝑎𝑙 𝑤𝑎𝑔𝑒 𝑝𝑟𝑒𝑚𝑖𝑢𝑚 ×

.

(A-4)

Using a perpetuity formula, equation (A-4) can be re-written as:
𝑃𝑉 𝑜𝑓 𝑤𝑎𝑔𝑒 𝑝𝑟𝑒𝑚𝑖𝑎
= 𝐴𝑛𝑛𝑢𝑎𝑙 𝑤𝑎𝑔𝑒 𝑝𝑟𝑒𝑚𝑖𝑢𝑚 + 𝑎𝑛𝑛𝑢𝑎𝑙 𝑤𝑎𝑔𝑒 𝑝𝑟𝑒𝑚𝑖𝑢𝑚 ×
=

× 1+

=

(A-5)

.

In general, the difference in the capitalized value of wage premia between any two firms
with risk-adjusted bankruptcy probabilities 𝑞 and 𝑞 (𝑞 > 𝑞 ) is equal to:
𝐷𝑖𝑓𝑓𝑒𝑟𝑒𝑛𝑐𝑒 𝑖𝑛 𝑃𝑉 𝑜𝑓 𝑤𝑎𝑔𝑒 𝑝𝑟𝑒𝑚𝑖𝑎 =

3

−

.

(A-6)

Note that this model does not require that the same employee work for the firm indefinitely. Workers receive a risk
premium to compensate them for the risk that the firm might go bankrupt in the current period. If the firm survives
the next period, it will compensate its workers for the risk of bankruptcy in that period, whether they are the same
employees that worked in the previous period or not.

4

For example, if a firm’s credit rating changes from AA to BBB, the capitalized value of
wage premia of AA over BBB ratings is equal to

−

.

A.3. Capitalized Value of Wage Premia, Adjusting for Pre-bankruptcy Wage Premium
Part of the present value of earnings loss L above might be due to the loss of a compensating
differential that existed pre-bankruptcy, which we attempt to tease out as follows. Suppose that the
average credit rating for the ultimately bankrupt firms is ‘B,’ while the average credit rating for
the control firms is ‘BBB,’ four years before bankruptcy. This assumption is based on the average
book leverage of the ultimately bankrupt and non-bankrupt firms (i.e., 41% and 27%), which very
closely map to ‘B’ and ‘BBB’ ratings in Table V of AP (2007) (i.e., 42% and 28%). Suppose that
the annual wage premium due to the difference in bankruptcy probability between BBB and Brated firms is 𝑊. This implies that the loss of the pre-bankruptcy wage premium contributes a
decline in earnings of W in each post-bankruptcy year from t to t+6. The present value of the loss
of this pre-bankruptcy wage premium from year t to t+6 is equal to W times the annuity factor for
seven years (𝑓), which is 𝑓𝑊. Therefore, the PV of “net” bankruptcy earnings losses is:
(A-7)

𝑁𝐿 = 𝐿 − 𝑓𝑊.

Given 𝑁𝐿, the worker will demand an annual wage premium (over a risk-free firm) that is equal
to

(

. Thus, the difference in annual wage premium between BBB and B-rated firms is:

)

𝑊=

−

Substituting (A-8) into (A-7), we get: 𝑁𝐿 = 𝐿 − 𝑓
we get:

5

.
−

(A-8)
. Solving the equation for NL,

𝑁𝐿 =

Given 𝑟 = 3.34%, 𝑓 = ∑

(

. %)

(

.

)

= 8.92,4 𝑞 = 1.51% and 𝑞

(A-9)

= 5.24% (from Table 8,

column (3)), the resulting value for NL relative to L suggests that the worker’s net earning loss due
to bankruptcy is about 76.2% of the “gross” loss.

4

-7.4% is the annual, risk-adjusted discount rate for the earnings loss conditional on a bankruptcy state, i.e., the
discount rate of earning losses from t to t+6. Online Appendix C provides details of how to derive this discount rate.

6

Online Appendix B
Certainty Equivalence Approach to Estimate Workers’ Risk-Neutral Probability of
Bankruptcy5
To estimate the difference in risk-neutral probabilities of the bankruptcy state between the worker
and bondholder, we use a certainty equivalence approach similar to that used by Hall and Murphy
(2002) who calculate the risk-adjusted value of option grants for under-diversified executives. We
first calculate the certainty equivalent values of labor income loss in the bankruptcy state for the
worker and bondholder. The worker and bondholder’s asset holdings are calibrated to match the
holdings of the average workers and bondholders in the Survey of Consumer Finance in 1998,
which is approximately in the middle of our sample period. We then calculate the relative
magnitudes of the worker and bondholder’s risk-neutral probabilities of bankruptcy using their
certainty equivalent values of labor income loss in the state. Details of the theoretical framework,
parameter calibration, and simulation results are described below.

B.1. Theoretical Framework
a) Worker
At Time 0, the worker has financial wealth 𝑊 allocated to directly held individual stocks,
indirectly held stocks through well diversified portfolios (such as equity mutual funds), the own
company’s stock, and the risk-free asset. We denote 𝑤

,𝑤

, and 𝑤

as the

proportions of the worker’s direct stock holdings, indirect stock holdings, and own company stock
in her financial assets, respectively. The proportion of the risk-free asset is equals to 𝑤

5

The simulation code for the certainty equivalence approach is available upon request.

7

=

1−𝑤

−𝑤

. 𝑁 is the number of directly held stocks. We assume that

−𝑤

the allocation to each individual stock in the direct holdings is equally weighted. In the normal
(i.e., non-bankruptcy) state, the worker will earn labor income 𝐼 . In the bankruptcy state, the
worker will lose a 𝜙 fraction of the labor income. The bankruptcy state is characterized by the
worker’s own company stock return going below a threshold level 𝑟̅ , i.e., 𝑟 < 𝑟̅ .
The worker’s wealth at time 1 is given by:
∑

𝑊 =𝑊
𝑤

𝑤

−𝑤

where 𝑟

1+𝑟

−𝑤
,

,

1+𝑟

(1 + 𝑟 ) + 𝑤

+𝑤

(1 + 𝑟 ) + 1 −
(A-10)

+ 𝐼 − 𝜙𝐼 ∙ 1(𝑟 < 𝑟̅ ),

is the return on directly held stocks i; 𝑟 is the return on indirectly held stocks; and

𝑟 is the return on the own company stock. 𝑟 is the risk-free rate. 1(𝑟 < 𝑟̅ ) is equal to one if 𝑟 <
𝑟̅ , and zero otherwise. To find the worker’s certainty equivalent for her uncertain labor income
loss 𝜙𝐼 ∙ 1(𝑟 < 𝑟̅ ), suppose that the worker is willing to give up 𝐶 in cash at Time 0 (which
she would have invested in the risk-free asset) in exchange for the uncertain labor income loss.
Her wealth at Time 1 would be:
𝑊

,

=𝑊

𝑤

∑

𝑤

−𝑤

1+𝑟

−𝑤

1+𝑟

,

(1 + 𝑟 ) + 𝑤

+𝑤

+𝐼 −𝐶

(1 + 𝑟 ) + 1 −
(A-11)

1+𝑟 .

Certainty equivalence requires that the worker achieves the same level of expected utility
with these two distributions of wealth at Time 1, i.e.,
𝐸𝑈(𝑊 ) = 𝐸𝑈 𝑊
∫ 𝑈(𝑊 )𝑓 𝑟
∫ 𝑈( 𝑊

,

) 𝑟

,

,..,𝑟
,

,..,𝑟

,𝑟 ,𝑟 𝑑 𝑟

,

,

,

,𝑟 ,𝑟 𝑑 𝑟

,..,𝑟
,

8

,

,..,𝑟

,𝑟 ,𝑟

,

,

=

,𝑟 ,𝑟 .

(A-12)

b) Bondholder
With a similar set-up, the bondholder’s wealth at Time 1 is given by:
∑

𝑊 =𝑊
𝑤

𝑤

−𝑤

1+𝑟

−𝑤

,

(1 + 𝑟 ) + 𝑤

+𝑤

(1 + 𝑟 ) + 1 −

+ 𝐼 − 𝜙𝐼 ∙ 1(𝑟 < 𝑟̅ ).

1+𝑟

(A-13)

The definitions of these variables are analogous to those for the worker except that the superscript
“b” indicates the variables are for the bondholder.
Similarly, to find the bondholder’s certainty equivalent for her uncertain labor income loss
𝜙𝐼 ∙ 1(𝑟 < 𝑟̅ ), suppose that the bondholder’s is willing to give up 𝐶 in cash at Time 0 (which
she would have invested in the risk-free asset) in exchange for the uncertain labor income loss.
Her wealth at Time 1 would be:
𝑊

,

=𝑊

𝑤

∑

𝑤

−𝑤

1+𝑟

−𝑤

1+𝑟

,

(1 + 𝑟 ) + 𝑤

+𝑤

(1 + 𝑟 ) + 1 −
(A-14)

+𝐼 −𝐶 1+𝑟 .

Certainty equivalence requires that the bondholder achieves the same level of expected
utility with these two distributions of wealth at Time 1, i.e.,
𝐸𝑈(𝑊 ) = 𝐸𝑈 𝑊
∫𝑈 𝑊

𝑓 𝑟

∫ 𝑈( 𝑊

,

) 𝑟

,

,..,𝑟
,

,..,𝑟

,𝑟 ,𝑟 𝑑 𝑟

,

,

,

,𝑟 ,𝑟 𝑑 𝑟

,

,..,𝑟
,

,..,𝑟

,𝑟 ,𝑟

,

,

=

,𝑟 ,𝑟 .

(A-15)

We solve equations (A-12) and (A-15) numerically to find a certainty equivalent loss for
the worker, 𝐶 and for the bondholder, 𝐶 . The implementation details are described in Sections
B.2 and B.3.
c) Converting Certainty Equivalent to Risk-Neutral Probability

9

The certainty equivalent of the labor income loss equals the risk-adjusted present value of
the loss, which implies:
𝐶 =

.

𝑜𝑟 𝑞 =

(A-16)

Similarly,
𝐶 =

𝑜𝑟 𝑞 =

Therefore,

=

/
/

.

(A-17)

. This result shows that if the fractional certainty equivalent of the labor

income loss is greater for the worker than for the bondholder (𝐶 /𝐼 > 𝐶 /𝐼 ), the risk-neutral
probability of bankruptcy is also greater for the worker than for the bondholder (

> 1). The

connection between the certainty equivalent and risk-neutral probability is intuitive; if the
bankruptcy state is more “painful” for the worker, she is willing to offer a greater certain loss in
exchange for the expected risky loss (i.e., certainty equivalent of the labor income loss in the
bankruptcy state is bigger). The more “painful” is the bankruptcy state, the higher the marginal
utility of consumption in (and thus the risk-neutral probability of) the bankruptcy state.
B.2. Parameters
To solve equations (A-12) and (A-15) numerically, we calibrate parameters of asset
allocations, the utility function, and the bankruptcy state.
a) Parameters of Asset Allocation and Labor Income
We collect detailed and realistic information on portfolio holdings of average workers with
financial assets and bondholders from the 1998 Survey of Consumer Finance (SCF). As the table
below shows, the average worker holds a considerably less diversified portfolio than the average
bondholder. For instance, the worker on average has four stocks in her directly owned stock
10

portfolio, whereas the bondholder has 11 stocks. Moreover, the worker holds more of her own
employer’s stock, which further increases her exposure to idiosyncratic risk.
Summary Statistics of Asset Allocations for Workers and Bondholders in 1998 SCF

Total financial assets (FA)
Age
Percentage own stocks
Indirectly held stocks/FA
Directly held stocks/FA
Number of directly held stocks, conditional on having
directly held stocks
Own company stocks/FA

Workers
$83,570
37.3
62.3%
25.1%
4.4%

Bondholders
$747,552
57.0
84.3%
28.7%
10.9%

4

11

4.1%

2.0%

Given the information from the 1998 SCF, the parameters of financial asset allocations are
set as follows:
Worker: 𝑊 = $83,570; 𝑁 = 4; 𝑤

= 4.4%; 𝑤

Bondholder: 𝑊 = $747,552; 𝑁 = 11; 𝑤

=10.9%; 𝑤

= 25.1%; 𝑤
= 28.7%; 𝑤

= 4.1%.
= 2%.

The worker’s labor income is set to the average labor income in our sample, i.e., 𝐿 =
$31,460; The bondholder’s labor income is set to the average employment earnings of bondholders
in the SCF, i.e., 𝐿 = $67,901.
b) Parameters of Risk-free Rate and Market Return


Nominal risk-free rate 𝑟 = 6.33% (the average 10-year Treasury yield during our sample
period of 1985-2008).



Expected market return 𝑟 = 11.35% (the average CRSP value-weighted annual return
(including dividends) from 1985 to 2008).



Standard deviation of market returns 𝜎 = 0.18 (the standard deviation of annual CRSP valueweighted return (including dividends) from 1985 to 2008).

11

c) Parameters of Directly Held Individual Stocks


Expected return of directly held individual stocks is the same as the expected market
return: 𝑟



,

=𝑟 .

Standard deviation of individual stock returns: 𝜎

,

= 0.58 (Fu, 2009).

d) Parameters of Indirectly Held Stocks
Indirectly held stocks are through mutual funds or pension funds which are well diversified,
so their expected return and variance are assumed to be the same as those of the market: 𝑟
𝑟 ; 𝜎

=

=𝜎 .

e) Parameters of Own Company Stock Return


Expected return of the own company stock is the same as the expected market return: 𝑟

,

=

𝑟 .


Standard deviation of the own company stock returns is the same as that of the directly held
stocks: 𝜎

=𝜎

,

.

f) Covariance between Stock Returns
The distribution of stock returns in Time 1 is lognormal with covariances between all stock
returns assumed to be the same and equal to the variance of the market return (0.18 = 1.32%).
Polkovnichenko (2003, 2005) note that this assumption is based on the fact that portfolio variance
converges to the weighted average covariance of stocks in the portfolio as the number of stocks
increases.6 It allows for imputation being internally consistent because the total variance of the

6

The variance of a portfolio 𝜎 (𝑅 ) = 𝐷𝑖𝑣𝑒𝑟𝑠𝑖𝑓𝑖𝑎𝑏𝑙𝑒 𝑟𝑖𝑠𝑘＋𝑁𝑜𝑛𝑑𝑖𝑣𝑒𝑟𝑠𝑖𝑓𝑖𝑎𝑏𝑙𝑒 𝑟𝑖𝑠𝑘 , where
𝐷𝑖𝑣𝑒𝑟𝑠𝑖𝑓𝑖𝑎𝑏𝑙𝑒 𝑟𝑖𝑠𝑘 = ∑ 𝑤 𝜎 (𝑟 ) , 𝑁𝑜𝑛𝑑𝑖𝑣𝑒𝑟𝑠𝑖𝑓𝑖𝑎𝑏𝑙𝑒 𝑟𝑖𝑠𝑘 = ∑ ∑ , 𝑤 𝑤 𝐶𝑜𝑣 𝑟 , 𝑟 ; 𝑤 and 𝑟 are the
weight and return of individual stock i, respectively. As the number of securities increases, 𝐷𝑖𝑣𝑒𝑟𝑠𝑖𝑓𝑖𝑎𝑏𝑙𝑒 𝑟𝑖𝑠𝑘
converges to 0, while 𝑁𝑜𝑛𝑑𝑖𝑣𝑒𝑟𝑠𝑖𝑓𝑖𝑎𝑏𝑙𝑒 𝑟𝑖𝑠𝑘 converges to 𝜎 when 𝐶𝑜𝑣 𝑟 , 𝑟 = 𝜎 .
12

stock portfolio of a household who owns a large number of different stocks would converge to the
market return variance.
g) Utility Function
The worker and bondholder are assumed to have a constant relative risk aversion utility
function 𝑈(𝑊) =

with relative risk aversion 𝛾 = 2 (Hall and Murphy, 2002).

h) Bankruptcy State
The worker’s (or bondholder’s) employer files for bankruptcy once the stock return falls
below a threshold that matches the historical bankruptcy probability by credit rating. Therefore,
for AAA-rated firms, the threshold 𝑟̅
𝑃𝑟𝑜𝑏(𝑟 < 𝑟̅

for the own company stock return is determined by

) = 0.0005, which is the historical bankruptcy probability for AAA-rated firms.

Similarly, thresholds for AA, A, BBB, BB and B-rated firms are determined by their historical
bankruptcy probabilities, which are equal to 0.0006, 0.0010, 0.0032, 0.0137, and 0.0322,
respectively.

B.3. Simulation
Given the above parameters, we first calculate the expected utility based on a simulation
of 1,000,000 sample stock returns. We then solve for the certainty equivalence relations in (A-12)
and (A-15) to find the risk-neutral probabilities for the worker and bondholder using (A-16) and
(A-17).
The table below shows the simulation results by credit rating (from ‘AAA’ to ‘B’).
Consistent with the worker’s greater exposure to idiosyncratic risk, which would make her require
a greater compensation for earnings loss risk in bankruptcy, the table shows that the worker’s risk-

13

adjusted bankruptcy probability (𝑞 ) is approximately 1.13 times (ranging between 1.11 and 1.14)
of the bondholder’s risk-adjusted probability (𝑞 ) across the ratings.
Table: Ratio between Risk-adjusted Bankruptcy Probabilities for Average Worker and
Bondholder
Ratio (𝑞 / 𝑞 )
1.11
1.11
1.12
1.14
1.14
1.14

Credit Rating
AAA
AA
A
BBB
BB
B

14

Online Appendix C
Estimating Workers’ Risk-Adjusted Discount Rate for Earnings Losses Given Bankruptcy
This appendix illustrates the approach to estimate the risk-adjusted discount rate for earnings losses
given corporate bankruptcy for workers who are exposed to considerable idiosyncratic risk. We
begin by building a simple model of the worker’s earnings process between year t (when the firm
files for bankruptcy) and year t+1. When the firm files for bankruptcy in year t, the worker’s annual
earnings become (1 − 𝜙)𝐼 , where 𝜙 is the average fractional earnings loss in bankruptcy and 𝐼
is the pre-bankruptcy annual earnings. In the following year t+1, there are two potential states: (1)
in a “good” state (with a probability 1-h), the worker’s earnings change by 𝑢, and (2) in a “bad”
state (with a probability h), the worker’s earnings change by 𝑑 (< u), both relative to earnings in
year t. Normalizing the pre-bankruptcy earnings (𝐼 ) to one, the situation between years t and t+1
is depicted by the following binomial tree.

Year t (Bankruptcy filing)

Year t+1
1-h

1−𝜙+𝑢
1−𝜙

1−𝜙+𝑑

h

+

For given 𝜙, h, and the variance of earnings growth between years t and t+1 (𝜎 ), u and d
should satisfy the following two conditions:
𝐸(𝑒𝑎𝑟𝑛𝑖𝑛𝑔𝑠 𝑔𝑟𝑜𝑤𝑡ℎ) = (1 − ℎ)(
𝑉𝑎𝑟(𝑒𝑎𝑟𝑛𝑖𝑛𝑔𝑠 𝑔𝑟𝑤𝑜𝑡ℎ) = (1 − ℎ)

) + ℎ(
+ ℎ(

15

) = 0,

(A-18)

) =𝜎 .

(A-19)

Solving these two equations, we get:
u=

∙(

)

d = −𝑢 ∙ (

).

Therefore, conditional on bankruptcy filing in year t, the present value (PV) of earnings loss in
year t+1 using risk-neutral probabilities equals:
𝑃𝑉(𝑙𝑜𝑠𝑠 ) =

=

(1 − 𝜋 )(−𝜙 + 𝑢) + 𝜋 (−𝜙 + 𝑑)
1+𝑟

(1 − 𝜋 )𝑢 + 𝜋 𝑑
−𝜙
+
,
1+𝑟
1+𝑟

where 𝜋 is the risk-neutral probability of state d for the worker and 𝑟 is the risk-free rate. The
first term represents the PV in case the loss is risk-free, and the second term is an adjustment for
the riskiness of the earnings loss in year t+1.
Using a risk-adjusted discount rate, the same present value of earnings loss in year t+1 is
equal to:
𝑃𝑉(𝑙𝑜𝑠𝑠 ) =

−𝜙
𝐸(𝑙𝑜𝑠𝑠 )
(1 − ℎ)(−𝜙 + 𝑢) + ℎ(−𝜙 + 𝑑)
=
=
,
1+𝑟
1+𝑟
1+𝑟

where 𝑟 is the risk-adjusted discount rate for the worker.
For the two approaches to produce the same risk-adjusted PV,
𝑃𝑉(𝑙𝑜𝑠𝑠 ) =

−𝜙 + (1 − 𝜋 )𝑢 + 𝜋 𝑑
−𝜙
=
,
1+𝑟
1+𝑟

or
𝑟

=

(1 + 𝑟 )𝜙
− 1.
𝜙 − [(1 − 𝜋 )𝑢 + 𝜋 𝑑]

The intuition of the above equation is the following:
a) If there is no correlation between a post-bankruptcy earnings loss and the marginal utility, 𝜋
is equal to h, the historical probability, so that 𝑟

= 𝑟 and we can discount the expected

earnings loss using the risk-free rate. (This is because if 𝜋
(1 − ℎ)𝑢 + ℎ𝑑 = 0 by equation (A-18)).
16

= ℎ, then (1 − 𝜋 )𝑢 + 𝜋 𝑑 =

b) If the post-bankruptcy earnings loss is positively correlated with the marginal utility, 𝜋
should be larger than h, which implies 𝑟

< 𝑟 . (This is because if 𝜋

> ℎ , then

(1 − 𝜋 )𝑢 + 𝜋 𝑑 < (1 − ℎ)𝑢 + ℎ𝑑 = 0 by equation (A-18)).
c) If the post-bankruptcy earnings loss is negatively correlated with the marginal utility, 𝜋
should be smaller than h, which implies 𝑟

> 𝑟 . (This is because if 𝜋

< ℎ , then

(1 − 𝜋 )𝑢 + 𝜋 𝑑 > (1 − ℎ)𝑢 + ℎ𝑑 = 0 by equation (A-18)).
To implement this approach, we make the following assumptions. First, the mean earnings
loss in year t (𝜙) is equal to 9.61% of pre-bankruptcy annual earnings, the average earnings loss
for each year from t through t+6 after UI payment adjustments (Table 4, Panel C). Second, without
loss of generality, we assume ℎ (probability of the “bad” state in year t+1, conditional on
bankruptcy filing in t) is equal to 0.5. Third, we follow Cocco, Gomes, and Maenhout (2005) and
Brown, Fang, and Gomes (2012), and assume 𝜎 (the standard deviation of annual earnings growth)
is equal to 0.10.7 These assumptions, combined with conditions in equations (A-18) and (A-19),
lead to u = 0.0904 and d = -0.0904. Lastly, we assume a real risk-free rate of 3.34%.
With these assumptions on the worker’s earnings risk between years t and t+1 conditional
on bankruptcy, we employ a certainty equivalence approach similar to that in Online Appendix B
(which is used to quantify the risk-adjusted probability of bankruptcy). While most of the
assumptions are the same (such as CRRA utility with a relative risk aversion of two), the key
differences include the definition of the two states and that labor income is assumed to be lower
by 9.61% for this simulation in year t (since we condition on bankruptcy). We find that 𝜋 , the
risk-adjusted probability of the “bad” state, is equal to 0.5614 > 0.5 = h, which in turn implies

7

Cocco, Gomes, and Meanhout (2005) use the PSID data and estimate the variance of permanent labor income shocks
to be 0.0106 (or an SD of 0.1030) and Brown, Fang, and Gomes (2012) estimate the variances of permanent labor
income shocks to be 0.0097 (or an SD of 0.0985) and 0.0100 (or an SD of 0.1000) for high school and college graduates,
respectively.

17

𝑟 = −7.4% < 𝑟 = 3.34%. This discount rate is smaller than the risk-free rate, consistent with
the worker’s relative lack of diversification, making her earnings losses in the bad state “more
painful.” We use this risk-adjusted discount rate to calculate the (risk-adjusted) PV of earnings
losses from years t through t+6.

18

Online Appendix D
Adjusting Earnings Losses for Unemployment Insurance Benefits
This appendix describes approaches to adjust the estimates of earnings losses after bankruptcy for
unemployment insurance (UI). For unemployed workers who satisfy the UI eligibility conditions
(see Section D.4), we calculate quarterly UI benefits using the following equation:
quarterly UI benefits = min(13 weeks × maximum weekly benefits, replacement rate × average
high-quarter(s) earnings during the base period before unemployment) × estimated UI take-up rate.
(A-20)
Below we provide the details on each variable in the equation. Table 4, Panel C and Online
Appendix Table A8 present the earnings loss estimates that account for all UI adjustments we
describe below.

D.1. Replacement Rate
The UI replacement rate is defined as the ratio of the UI claimants’ weekly benefit amount
to the claimants’ average weekly wage. We collect the state-year-level unemployment insurance
benefits schedule from the “Significant Provisions of State Unemployment Insurance Laws” (the
“Provisions” hereafter) issued by the U.S. Department of Labor (DOL) over our sample period
and estimate the replacement rate for each state and year. 8 The Provisions contain information on
the calculation of the weekly benefit amount in each state and each year. The weekly benefit
amount is generally specified as a percentage of employees’ representative wages before
unemployment. For example, the weekly benefit amount is 1/26 of high-quarter wages for Florida
in 2000, where the “high quarter” is the quarter with the highest wages among the base period,

8

The Provisions are obtained from https://oui.doleta.gov/unemploy/laws.asp.

19

defined as the first four of the last five completed calendar quarters. We record the UI replacement
rate as 50% (= 1/26 × 13 weeks) for this case. When the Provisions give a range of the weekly
benefit amount, we use a mid-point value. For example, the weekly benefit is 1/23 to 1/33 of the
high-quarter wage in California in 2000, so we use 46.43% (= 1/28 × 13 weeks) as the (estimated)
replacement rate.

D.2. High-quarter Wages
By following the Provisions for each state and year, we separately code the earnings that
UI benefits are based on (“high-quarter wages”). We find that for about two-thirds of our sample
state-years, the specified weekly benefits are based on the highest-quarter wages during the base
period. For the rest of the approximately one-third sample state-years, UI benefits are based on the
average of highest two through four quarters in the base period. For these state-years, we revise
our formula accordingly.

D.3. Maximum Benefit Amount
The Provisions also provide information on the maximum weekly benefit amount. For
example, the Provisions state that the maximum weekly benefit is $435 in Colorado in 2007. We
use this information to cap the amount of UI benefits received by each unemployed worker.

D.4. Take-up Rate
We note that not every unemployed worker applies for and receives the UI even if he or
she is eligible. We thus need to calculate a UI take-up rate, defined as the proportion of the
unemployed receiving UI out of all UI-eligible unemployed. Because UI take-up rates are not
20

directly available from public sources such as US DOL publications (e.g., Auray et al., 2019), we
estimate the state-year-level take-up rates by combining data from the US DOL and LEHD and
following the procedures provided in Auray et al. (2019). First, we write the UI take-up rate as
follows:
Take up rate = recipiency rate/fraction eligible,

(A-21)

where the recipiency rate is the number of unemployed individuals that have collected or are
collecting UI benefits divided by the total number of unemployed individuals, and the fraction
eligible is the number of unemployed individuals who are eligible to collect UI benefits divided
by the total number of unemployed individuals.
We obtain the UI recipiency rate by state and year from the US DOL website. 9 The data
show that UI recipiency rates vary across states. This suggests that incorporating variation in UI
recipiency rates across states is important for accurately calculating UI benefits.
We then use the worker-level LEHD data to estimate the fraction eligible for UI (second
term in equation (A-21)). Eligibility depends primarily on three factors: (1) monetary requirements;
(2) the duration of unemployment; and (3) the nature of separation, which should not be voluntary
(see e.g., Blank and Card (1991)). (1) Monetary requirements are obtained from the Provisions.
For each state and year, the Provisions provide information on the employment earnings needed
in the base period for workers to qualify for UI. For example, the Provision for Florida in January
2002 specifies that the base-period earnings required to qualify for UI is the maximum of $3,400
and 1.5 times the worker’s highest-quarter wage. (2) The maximum number of UI benefit weeks
is 26 weeks for all sample states (except for Montana which has 28 weeks). If an individual has

9

https://oui.doleta.gov/unemploy/data_summary/DataSum.asp, then choose “Labor Force Data.”

21

collected benefits for 26 (or 28) total weeks but remained unemployed, they are no longer eligible
to collect, unless extended benefits have been triggered (see Section D.5).
For (3), we follow Anderson and Meyer (1997) to exclude spurious transitions and
voluntary quits. We assume that separation cases in which an individual has no reported earnings
in one or two quarters, but earnings before and after this gap are nearly identical (i.e., within 5%
of each other) are not actually unemployed and thus not eligible for UI, since these cases are likely
due to a firm neglecting to report wages, as they argue. As a robustness test, we impute these
missing quarterly earnings that firms appear to neglect to report using the average of quarterly
earnings right before and right after the gap. The PV earnings loss estimate (Online Appendix
Table A5, column (7)) shows that the magnitude (87.9%) is similar to our baseline estimate
(89.5%). Anderson and Meyer (1997) also exclude separations that result in only a slight drop in
usual earnings (i.e., less than 2/13 of quarterly earnings, equivalent to under two weeks’
unemployment) because such separations are likely to be quits related to a quick job move rather
than unemployment. This type of separation is not captured as unemployment in our definitions,
since we define unemployment as either having zero quarterly earnings or more than a half drop
in quarterly earnings.
We note that UI take-up may be higher among individuals in longer unemployment spells,
and thus adjust the state-year level take-up rate (recipiency rate in particular) for unemployment
durations as follows. We follow Vroman (2009) and use the 2005 Current Population Survey Unemployment Insurance Supplemental Survey (2005 UI survey) to estimate the UI recipiency
rate by group of unemployment spell length. 10 The groups include: a) less than a quarter of

10

According to Vroman (2009), Current Population Survey – Unemployment Insurance Supplemental Surveys were
conducted in 1976, 1989, and 1993, and 2005. Since 2006, the survey was only conducted in 2018. We contacted
Vroman and Research Director of Department of Labor, and were advised that the 1976, 1989, and 1993 survey data

22

unemployment (which is captured by reduced yet positive quarterly earnings); b) a quarter of
unemployment (captured by a full quarter of zero or missing earnings); c) between a quarter and
two quarters of unemployment; d) two quarters of unemployment; and e) more than two quarters
of unemployment.
We then apply a “premium” or “discount” to the state-year-level average recipiency rate,
conditional on the length of unemployment spells. For example, the UI survey data show that the
recipiency rate for those with 14-26 weeks of unemployment is 41% higher than the average
recipiency rate. Then, for Florida in 2005-Q3, in which the UI recipiency rate is 27.9%, we
estimate the recipiency rate for those with 14-26 weeks of unemployment duration to be 27.9% ×
1.41 = 39.3%.
Finally, combing the duration-dependent recipiency rate and the average fraction of UI
eligible workers, we obtain the UI take-up rates using equation (A-21) for different unemployment
durations. We use the resulting UI take-up rate by state-year-duration group in our UI benefits
calculations. The average take-up rate across all unemployed workers in our sample is 72.6% and
is in line with existing estimates (e.g., 70.7% over 1977-1987 from Blank and Card (1991)).

D.5. Extended UI Benefits
UI benefits include not only the regular benefits as specified in the Provisions, but also
extended benefits, which consist of the permanent federal-state extended benefits (EB) and
temporary emergency unemployment compensation (EUC) programs. The permanent EB program
provides an additional 13 weeks of benefits to workers who have exhausted regular UI benefits

are currently not available for researchers to use. Therefore, we use the 2005 survey data, which is within our sample
period, to estimate the recipiency rate by duration of unemployment. We do not use the 2018 data because our sample
ends well before 2018.

23

when a state meets a high unemployment rate threshold. For states with extremely high
unemployment rates, the benefits are extended for 20 weeks (see e.g., Hsu, Matsa, and Melzer,
2018). We use the “Extended Benefit Trigger Notice Reports” to obtain the information on the
start and end dates for the EB program by state and whether it is for 13 or 20 additional weeks. 11
In addition to the EB program, we collect information on the three temporary EUC
programs that are applicable to our sample period from 1990 to 2008: EUC (Emergency
Unemployment Compensation) 1991-94, TEUC (Temporary Extended Unemployment
Compensation) 2002-04, and EUC (Emergency Unemployment Compensation) 2008. EUC 199194, effective from November 1991 to April 1994 and applicable to all states, covers individuals
who exhausted their regular and extended benefits. Additionally, the benefits received under the
EUC program are reduced by any EB received (Lake, 2002). The maximum number of benefit
weeks under EUC 1991-94 is based on unemployment rates of states. The maximum number of
benefit weeks also varies over periods, ranging from 13 to 33 weeks. For example, the maximum
number of benefit weeks across all states was 26 from June 14, 1992 to March 6, 1993 and 15 for
the period from March 7, 1993 to October 2, 1993. We obtain the EUC 1991-94 benefit duration
in weeks by state from Table B of Lake (2002).
The second emergency unemployment insurance program, TEUC 2002-04, was effective
from March 2002 to March 2004. This program includes two tiers; the first tier provides up to 13
weeks of benefits for all states, and the second tier, TEUC-X, provides up to 26 weeks in the high-

11

The trigger notice reports starting in October 2002 are available on the website
https://oui.doleta.gov/unemploy/claims_arch.asp. Based on the laws on EB, trigger notices should be provided to the
public via the Federal Register. For the period between 1995 and 2002, we collect the “Notice of Changes in Status
of Extended Benefit (EB) Periods” on www.federalregister.gov by searching for “extended benefit” as keywords. For
years before 1995, a digitized federal register in scanned newspaper format is obtained from www.gpo.gov and the
information on EB is obtained through searching using keywords including “federal-state unemployment
compensation program” and “extended benefit.”

24

unemployment states (Whittaker and Isaacs (2014b)). 12 In contrast to EUC 1991-94, TEUC 200204 supplements the EB program. When a state has the EB program triggered during the same
period, the benefit weeks from the TEUC program are added to regular UI and EB weeks.
The third program, EUC 2008, was in effect from July 2008 till December 2013. Since our
sample ends in 2008, EUC 2008 is applicable only up to December 2008. EUC 2008 provides up
to 13 weeks of benefits for all states from July 6, 2008 till November 22, 2008. From November
23, 2008 to March 28, 2009, there are two tiers under EUC 2008. Tier 1 provides up to 20 benefit
weeks for all states, and tier 2 provides 13 additional weeks (33 weeks in total) if the state
unemployment rate is higher than a pre-specified threshold (i.e., state total unemployment rate
(TUR) is 6% or higher, or insured unemployed rate (IUR) is 4% or higher) (Table A-4 of Whittaker
and Isaacs (2014b)). 13 According to Whittaker and Isaacs (2014a), EUC 2008 benefits are in
addition to regular UI and EB. However, states pay EUC 2008 benefits after regular UI exhaustion,
but before EB benefits prior to EUC 2008 expiration.
To summarize, we include the permanent EB and three different temporary EUC programs,
in addition to regular UI benefits in our analysis. The regular UI provides up to 26 weeks of
benefits (28 weeks for Montana), the EB program up to 20 weeks, EUC 1991-94 up to 33 weeks,
TEUC 2002-04 a maximum of 26 weeks, and EUC 2008 up to 33 weeks until March 2009. In
general, when the EUC programs are not in effect, a worker could have a maximum of 46 (= 26 +
20) weeks of benefits. When EUC 1991-94 was in effect, a worker’s maximum number of benefit

12

The high-unemployment states that trigger TEUC-X are obtained from the trigger notices at
https://oui.doleta.gov/unemploy/teuc/2002/teuc_101302.html.
13
We use the trigger notices on the DOL website (https://oui.doleta.gov/unemploy/claims_arch.asp) to find the tiers
of benefits each state qualifies for within the EUC 2008 program.

25

weeks was 59 (= 26 + 33). The maximum number of weeks was 72 (= 26 + 20 + 26) when TEUC
2002-04 was in effect, and 79 (= 26 + 20 + 33) when EUC 2008 was in effect.
Finally, to include these extended benefits in our calculation of UI benefits, we first infer
a worker’s duration of unemployment from her quarterly earnings. If the duration is greater than
26 weeks (the maximum number of weeks one is typically entitled to regular UI), then we check
whether there is any active EB or EUC program during the unemployment spell. If there are, then
we add the additional benefit weeks to the regular (typically) 26 weeks. It is possible that the
maximum available EB and EUC benefit weeks change during a worker’s unemployment spell.
For example, the EB duration could increase from 13 to 20 weeks in a very high unemployment
period. In this case, we choose the maximum (i.e., 20 weeks of EB in the example above). Of
course we require that the total number of benefit weeks not exceed the worker’s unemployment
duration.

D.6. Accounting for Short Unemployment Spells
When a worker is laid off at the beginning of one quarter and reemployed during the same
quarter, the unemployment spell for the worker spans less than an entire quarter and the worker
may still have positive earnings during this quarter. Our main analysis may thus underestimate the
amount of UI received by the worker because we attribute this quarter with positive earnings as
employed and ignore UI received during this “employed” quarter. Similarly, even some
unemployment spells lasting more than a quarter (e.g., 5.5 months) can be missed if the worker is
laid off at the beginning of one quarter and reemployed at the end of the subsequent quarter.
To account for these missing spells, we assume that a temporary drop in wages following
bankruptcy (or matching) is attributable to unemployment. Specifically, we assume that an
26

unemployment spell begins when quarterly earnings drop by at least half of the previous quarterly
earnings or becomes zero, the latter being our definition in the main analysis. In addition, we
assume that an unemployment spell may end with positive quarterly earnings that are lower than
the pre-unemployment quarterly earnings, in which case the worker can receive UI payments.
These adjustments would fully capture UI benefits during a 5.5-month unemployment spell across
two calendar quarters, for example.
The results for this bounding exercise show that in our preferred specification, the present
value (PV) of earnings losses with this alternative UI adjustment (84.2%) is slightly smaller in
magnitude than the PV with our main UI adjustment (87.1%). Therefore, this alternative definition
of unemployment allows workers to be unemployed and eligible for UI benefits more frequently
and potentially longer. Given that the threshold to define unemployment with positive quarterly
earnings is arbitrary and that the economic magnitudes in both UI adjustments are similar, we
present the PV with the alternative UI adjustment as a robustness check in Online Appendix Table
A5, column (3).

27

References:
Almeida, Heitor and Thomas Philippon, 2007, The Risk-Adjusted Cost of Financial Distress,
Journal of Finance 62, 2557–2586.
Anderson, Patricia and Bruce Meyer, 1997, Unemployment Insurance Takeup Rates and the AfterTax Value of Benefits, Quarterly Journal of Economics 112, 913–937.
Auray, Stephane, David L. Fuller, and Damba Lkhagvasuren, 2019, Unemployment Insurance
Take-up Rates in an Equilibrium Search Model, European Economic Review 112, 1–31.
Blank, Rebecca M. and David E. Card, 1991, Recent Trends in Insured and Uninsured
Unemployment: Is There an Explanation?, Quarterly Journal of Economics 106, 1157-1189.
Brown, Jeffrey, Chichun Fang, and Francisco J. Gomes, 2012, Risk and Returns to Education,
NBER Working paper 18300.
Cocco, João F., Francisco J. Gomes, and Pascal J. Maenhout, 2005, Consumption and Portfolio
Choice over the Life Cycle, Review of Financial Studies 18, 491–533.
Hall, Brian J. and Kevin Murphy, 2002, Stock Options for Undiversified Executives, Journal of
Accounting and Economics 33, 3–42.
Hsu Joanne W., David A. Matsa, and Brian T. Melzer, 2018, Unemployment Insurance as a
Housing Market Stabilizer, American Economic Review 108, 49–81.
Lake, Jennifer E., 2002, Temporary Programs to Extend Unemployment Compensation, CRS
(Congressional

Research

Service)

Report

for

Congress

(available

at

https://www.everycrsreport.com/files/20020326_RL31277_0ab8f9eae7b66dc727e8a4096e7a
ad5a65c646b8.pdf).
Polkovnichenko, Valery, 2003, Household Portfolio Diversification, Working paper, Federal
Reserve Bank of Minneapolis.
Polkovnichenko, Valery, 2005, Household Portfolio Diversification: A Case for Rank-Dependent
Preferences, Review of Financial Studies 18, 1467–1502.
Vroman, Wayne, 2009, Unemployment Insurance Recipients and Nonrecipients in the CPS,
Monthly Labor Review (available at https://www.bls.gov/opub/mlr/2009/10/art4full.pdf).
Whittaker, Julie M. and Katelin P. Isaacs, 2014a, Unemployment Insurance: Programs and
Benefits, CRS (Congressional Research Service) Report for Congress (available at
https://fas.org/sgp/crs/misc/RL33362.pdf).
28

Whittaker, Julie M. and Katelin P. Isaacs, 2014b, Extending Unemployment Compensation
Benefits During Recessions, CRS (Congressional Research Service) Report for Congress
(available at https://crsreports.congress.gov/product/pdf/RL/RL34340).

29

Online Appendix Figure A1: Dynamics of Firm Characteristics for Bankrupt and Matched
Firms Prior to Bankruptcy
This figure plots the descriptive statistics on dynamics of firm characteristics for bankrupt and matched firms from
years t-3 to t-1, which are reported in Table 2, Panel C in the paper.

Sales ($m)

Book value of assets ($m)
1,600

2,000

1,500

1,800

1,400

1,600

1,300

1,400

1,200

1,200

1,100
1,000

1,000
t-3

t-2

Bankrupt firms

t-3

t-1
Matched firms

t-2

Bankrupt firms

t-1
Matched firms

Log market value of assets

Market value of assets ($m)
6.7
6.6
6.5
6.4
6.3
6.2
6.1
6.0

1,700
1,500
1,300
1,100
900
t-3

t-2

t-3

t-1

Bankrupt firms

t-2

Bankrupt Firms

Matched firms

Δlog market value of assets
t-2

Matched Firms

N. Emp. (LBD)

0.00
t-3

t-1

15,000

t-1

-0.05

13,000

-0.10

11,000

-0.15

9,000
7,000

-0.20

5,000

-0.25

3,000
-0.30

t-3
Bankrupt Firms

Matched Firms

Bankrupt firms

30

t-2

t-1
Matched firms

Book leverage

Market leverage

0.80

0.80

0.70

0.70

0.60

0.60

0.50

0.50
0.40

0.40

0.30

0.30

0.20

0.20

0.10
t-3

t-2

Bankrupt firms

t-1

t-3

Matched firms

t-2

Bankrupt firms

t-1
Matched firms

Market-to-book
1.80

ROA

1.70

0.16

1.60

0.14

1.50

0.12

1.40

0.10
0.08

1.30

0.06

1.20

0.04

1.10
1.00

0.02
t-3
Bankrupt firms

t-2

t-3

t-1

Bankrupt firms

Matched firms

31

t-2

t-1
Matched firms

Wages/Assets

Capex/Net PP&E

0.40

0.50

0.35

0.45

0.30

0.40
0.35

0.25

0.30

0.20

0.25
0.20

0.15

0.15

0.10

0.10
t-3

t-2

Bankrupt firms

t-1

t-3

Matched firms

t-2

Bankrupt firms

R&D/Assets
0.020
0.015
0.010
0.005
0.000
t-3

t-2

Bankrupt firms

32

t-1
Matched firms

t-1
Matched firms

Online Appendix Figure A2: Dynamics of Employee Earnings after Corporate Bankruptcy
- Conditional Analysis on Bankruptcy Liquidation Outcomes
This figure plots difference-in-differences estimates of earnings changes for workers employed by bankrupt firms
around bankruptcy filings relative to the matched control group of workers, conditional on bankruptcy liquidation
outcomes. Panel A (Panel B) shows the estimates for bankruptcy events that ultimately lead to liquidation (nonliquidation). The present values of the earnings loss estimates are reported in Table 5, Panel B in the paper.

Panel A: Liquidation
0.30

Log earnings change

0.20
0.10
0.00
t

t+1

t+2

t+3

t+4

t+5

t+6

t+5

t+6

-0.10
-0.20
-0.30
-0.40
-0.50

Years relative to bankruptcy

Panel B: Non-liquidation
0.30

Log earnings change

0.20
0.10
0.00
t

t+1

t+2

t+3

t+4

-0.10
-0.20
-0.30
-0.40
-0.50

Years relative to bankruptcy

33

Online Appendix Figure A3: Covenant Violations and Employee Earnings Changes - An
RDD Graphical Analysis
This figure plots the average cumulative growth in a worker’s log(annual earnings) in 2001 constant dollars over years
t and t+1 relative to the average real log(annual earnings) in years t-1 and t-2 (y-axis) against the relative average
distance from the net worth and current ratio covenant thresholds (x-axis). We employ the cumulative earnings growth
and average distance given the relatively small firm-year sample using the DealScan-Compustat-LEHD matched data
from 1995 to 2007. Each circle represents the average cumulative earnings growth within the bin, with each bin
containing multiple underlying worker observations. Solid lines are fitted values from kernel-weighted polynomial
regressions and the dashed lines represent the 95% confidence intervals.

34

Online Appendix Table A1: Effect of Bankruptcy on Employee Earnings When
Control Group is Matched Using Observations from the Baseline Year, t-4
This table presents the present value of earnings losses by constructing matched firms based on the information in the
baseline year, t-4. “PV (Earnings losses)” is the present value of earnings losses from years t to t+6 as a percent of the
pre-bankruptcy annual earnings in t-4. Heteroskedasticity robust t-statistics adjusted for within-firm and worker
clustering are in parentheses (standard errors are computed using the delta method). Statistical significance at the 10%,
5%, and 1% levels is indicated by *, **, and ***, respectively. The number of observations is rounded to the nearest
thousands to follow the Census Bureau’s disclosure rules.

PV (Earnings losses)
Controls
Worker FE
Year FE
SIC2 × Year FE
County × Year FE
SIC2 × County × Year FE
Observations

(1)
-153.5%***
(-4.24)
Yes
Yes
Yes

(2)
-163.1%***
(-4.44)
Yes
Yes

(3)
-66.4%*
(-1.68)
Yes
Yes

(4)
-128.8%***
(-3.82)
Yes
Yes

Yes

Yes
Yes

2,151,000

2,151,000

Yes

2,151,000

2,151,000

35

(5)
-148.3%**
(-2.43)
Yes
Yes

Yes
2,151,000

Online Appendix Table A2: Effect of Corporate Bankruptcy on Employee Earnings After
Imputing Missing Earnings
This table presents the present value of earnings losses using the matched firms’ employees as the control group and
after imputing missing earnings in LEHD. “PV (Earnings losses)” (“PV (Earnings losses, UI)”) is the present value
of earnings losses without (with) unemployment insurance adjustments from years t to t+6 as a percent of the prebankruptcy annual earnings in t-4. Column (1) shows the baseline PV estimates from Table 4, Panels A and C, column
(5). Columns (2) through (4) impute missing earnings in LEHD with the 0.5%, 1%, and 5% values of the earnings
distribution in our sample. Columns (5) and (6) impute missing earnings in LEHD with the last observed quarterly
and annual earnings for a given worker. Heteroskedasticity robust t-statistics adjusted for within-firm and worker
clustering are in parentheses (standard errors are computed using the delta method). Statistical significance at the 10%,
5%, and 1% levels is indicated by *, **, and ***, respectively.

PV (Earnings losses)
PV (Earnings losses, UI)
Controls
Worker FE
SIC2 × County × Year FE

(1)
Baseline

(2)
Impute with
0.5% earnings

(3)
Impute with
1% earnings

(4)
Impute with
5% earnings

-89.5%**
(-2.14)
-87.1%**
(-2.07)
Yes
Yes
Yes

-169.3%**
(-2.22)
-115.6%***
(-2.70)
Yes
Yes
Yes

-157.5%**
(-2.29)
-113.9%***
(-2.88)
Yes
Yes
Yes

-121.4%**
(-2.55)
-107.5%***
(-3.36)
Yes
Yes
Yes

36

(5)
Impute with
last observed
quarterly
earnings
-63.6%**
(-2.44)
-62.2%**
(-2.35)
Yes
Yes
Yes

(6)
Impute with
last observed
annual earnings
-83.5%***
(-3.17)
-82.2%***
(-3.10)
Yes
Yes
Yes

Online Appendix Table A3: Alternative Definition of Firm, Industry, and County Stayers
in Year t+3 vs. t+5
This table presents the present value of earnings losses conditional on mobility of employees of the bankrupt firms.
“PV (Earnings losses)” is the present value of earnings losses from years t to t+6 as a percent of the pre-bankruptcy
annual earnings in t-4. Panel A (Panel B) defines a worker as a “stayer” if she stays in the same firm, industry, or
county through year t+3 (t+5) as her t-1 firm, industry, or county. Heteroskedasticity robust t-statistics adjusted for
within-firm and worker clustering are in parentheses (standard errors are computed using the delta method). Statistical
significance at the 10%, 5%, and 1% levels is indicated by *, **, and ***, respectively.

(1)
Group:

Firm stayers

(2)

(3)
(4)
Leavers:
Leavers:
Firm leavers
same industry diff. industry
same county
same county
Panel A: Stayers Defined in t+3

(5)
Leavers:
same industry
diff. county

(6)
Leavers:
diff. industry
diff. county

PV (Earnings
losses)

37.9%
(0.62)

-119.9%***
79.1%
19.0%
(-3.11)
(1.42)
(0.39)
Panel B: Stayers Defined in t+5

8.5%
(0.16)

-211.1%***
(-6.02)

PV (Earnings
losses)
Controls
Worker FE
SIC2 × County
× Year FE

107.0%
(1.45)
Yes
Yes

-110.7%***
(-2.77)
Yes
Yes

130.7%**
(2.17)
Yes
Yes

28.9%
(0.57)
Yes
Yes

22.2%
(0.41)
Yes
Yes

-187.6%***
(-5.07)
Yes
Yes

Yes

Yes

Yes

Yes

Yes

Yes

37

Online Appendix Table A4: Present Value of Earnings Losses Conditional on Firm,
Industry, and County Stayer Status of both Treated and Control Groups
This table presents the present value of earnings losses conditional on mobility of employees of both bankrupt and
matched control firms. The main difference between this table and Table 5, Panel A in the main text is as follows.
This table stratifies both the treated (i.e., bankrupt) and control firm employees. Thus, this table examines how stayers
(leavers) of the bankrupt firm fare relative to the stayers (leavers) of the control firm. In contrast, Table 5, Panel A
stratifies the treated firm employees based on whether they stay with their firms, industries, or counties, but not the
control firm employees. Thus, the panel decomposes earnings losses of the treated group relative to the average worker
in the control group. “PV (Earnings losses)” is the present value of earnings losses from years t to t+6 as a percent of
the pre-bankruptcy annual earnings in t-4. We define a worker as a stayer if she stays in the same firm, industry, or
county through year t+3 as her t-1 firm, industry, or county. Heteroskedasticity robust t-statistics adjusted for withinfirm and worker clustering are in parentheses (standard errors are computed using the delta method). Statistical
significance at the 10%, 5%, and 1% levels is indicated by *, **, and ***, respectively.

(1)

(2)

-86.8%**
(-2.59)
Yes
Yes

(3)
Leavers:
same industry
same county
-97.0%**
(-2.23)
Yes
Yes

(4)
Leavers:
diff. industry
same county
-94.9%**
(-2.41)
Yes
Yes

(5)
Leavers:
same industry
diff. county
-143.7%***
(-3.56)
Yes
Yes

(6)
Leavers:
diff. industry
diff. county
-106.5%***
(-3.14)
Yes
Yes

Group:

Firm stayers

Firm leavers

PV (Earnings
losses)
Controls
Worker FE
SIC2 × County
× Year FE

-116.2%**
(-2.00)
Yes
Yes

Yes

Yes

Yes

Yes

Yes

Yes

38

Online Appendix Table A5: Present Value of Earnings Losses Estimates Using Alternative Samples and Specifications
This table presents the present value of earnings losses using various alternative samples and specifications to the baseline results. Column (1) examines the
earnings losses of employees of private bankrupt firms relative to matched private non-bankrupt firms; Column (2) focuses on a subsample of employees in the
treated and control groups of the main sample with at least six years of tenure in year t-1; Column (3) estimates equation (1) using the unemployment insurance
adjusted earnings, in which unemployment spells are defined to begin when quarterly earnings either drop by at least half of the previous quarterly earnings or
become zero; Column (4) focuses on a subsample of firms that have existed in years t-3 through t-1, corresponding to the subsample used in Table 2, Panel C;
Column (5) controls for worker education by using five categories of education level, instead of the number of schooling years in the main specification; Column
(6) uses employees of bankrupt and matched firms with at least two years of tenure in year t-4 as a sample; Column (7) imputes one or two quarters of zero
earnings that are preceded and followed by nearly identical positive quarterly earnings (i.e., within 5% of each other) with the average of the positive earnings.
Statistical significance at the 10%, 5%, and 1% levels is indicated by *, **, and ***. The number of observations is rounded to the nearest thousands to follow the
Census Bureau’s disclosure rules.

Description:
PV (Earnings losses)
Controls
Worker FE
SIC2 × County × Year FE
Observations

(1)

(2)

(3)

(4)

(5)

(6)

(7)

Private firms
-289.3%***
(-6.08)
Yes
Yes
Yes

Tenure ≥ 6
-84.8%
(-1.34)
Yes
Yes
Yes

Unemp. w/
pos. earnings
-84.2%**
(-2.04)
Yes
Yes
Yes

Firms existed
in t-3 to t-1
-133.9%***
(-2.81)
Yes
Yes
Yes

Discrete
education
-89.1%**
(-2.13)
Yes
Yes
Yes

Workers in
year t-4
-62.0%***
(-2.70)
Yes
Yes
Yes

332,000

781,000

2,967,000

2,588,000

2,967,000

1,770,000

Spurious zeros
imputed
-87.9%**
(-2.12)
Yes
Yes
Yes
2,967,000

39

Online Appendix Table A6: Covariate Balance Before Potential Loan Covenant Violations
This table employ a regression discontinuity design to show the balance of worker characteristics around covenant
violation thresholds using DealScan-Compustat-LEHD matched data from 1995 to 2008. The dependent variable is
the worker’s annual earnings in 2001 constant dollars in the year before potential violations (column (1)), an indicator
for female (column (2)), years of experience (column (3)), imputed education (column (4)), and age (column (5)).
“Bind” is an indicator variable equal to one if the worker is an employee of a firm that violates a loan covenant
concerning the current ratio or net worth, and zero otherwise. “1 st order distance” is the relative distance of the firm’s
net worth and current ratio from the covenant thresholds. Definitions of these variables are available in the Appendix.
Heteroskedasticity robust t-statistics adjusted for within-firm and worker clustering are in parentheses. Statistical
significance at the 10%, 5%, and 1% levels is indicated by *, **, and ***. The number of observations is rounded to the
nearest thousands to follow the Census Bureau’s disclosure rules.

Dep. Var.:
Bind
1st order distance
Worker-level controls
Firm FE
SIC2 × County × Year FE
Observations
R2

(1)
Earnings ($)
-162.3
(-0.22)

(2)
Female
-0.002
(-0.49)

(3)
Experience
-0.040
(-0.42)

(4)
Education
-0.026
(-1.21)

(5)
Age
-0.066
(-0.64)

Yes

Yes

Yes

Yes

Yes

Yes
Yes
Yes
1,673,000
8.04%

Yes
Yes
Yes
1,673,000
21.46%

Yes
Yes
Yes
1,673,000
15.89%

Yes
Yes
Yes
1,673,000
7.03%

Yes
Yes
Yes
1,673,000
17.42%

40

Online Appendix Table A7: Leverage and Employee Wages - A Wage Regression
Approach Using Alternative Samples
This table presents the relationship between corporate leverage and employee earnings using alternative samples to
those used in Table 7. In Panel A, the dependent variable is log(quarterly earnings) in 2001 constant dollars. The
control variables include log(book assets), market-to-book, ROA, asset tangibility of firms, female indicator × years
of experience, imputed education × years of experience of individual workers, firm, worker, and two-digit SIC industry
× county × year fixed effects. To select the sample for the regression, we begin with all Compustat firms, except for
those with S&P credit ratings D (default), with book assets greater than $100 million in 1980 constant dollars and less
than or equal to the assets of the largest firm in our bankrupt and matched firm sample from 1986 to 2008. Then, we
select all workers in LEHD with ages between 22 and 55 with at least two years of tenure who move between these
firms. For earnings before moving, we use the highest quarterly earnings in the last four quarters at the old firm; for
earnings after moving, we use the highest quarterly earnings in the first four quarters at the new firm. Column (1)
presents estimates for the full sample. Columns (2) and (3) present estimates conditional on firm size, as measured by
book assets of a given firm. Columns (4) and (5) present estimates conditional on local labor market size, as measured
by total employment in a given two-digit SIC industry and county cell. “Large” (“Small”) represents a subsample in
which firm assets or local labor market size is greater than (less than or equal to) the 75 th percentile in the sample
distribution. In Panel B, the dependent variable is log(annual earnings) in 2001 constant dollars. To select the sample
for the regression, we begin with Compustat firms with S&P credit ratings between AA+ and B- or unrated, and book
assets greater than $100 million in 1980 constant dollars and less than or equal to the assets of the largest firm in our
bankrupt and matched firm sample from 1986 to 2008. Then we draw a 30% random sample of all employees of these
firms in LEHD (to reduce computational burden) with ages between 22 and 55 and at least two years of tenure. The
same sets of worker- and firm-level controls as those in Panel A are employed. Definitions of these variables are in
the Appendix. Heteroskedasticity robust t-statistics adjusted for within-firm and worker clustering are in parentheses.
Statistical significance at the 10%, 5%, and 1% levels is indicated by *, **, and ***. The number of observations is
rounded to the nearest thousands to follow the Census Bureau’s disclosure rules.

Panel A: Mover Sample Including All Rated and Unrated Firms
Dep. Var. = Log (earnings)
Conditioning variable:
Sample:
Market leverage
Worker-level controls
Firm-level controls
Worker FE
Firm FE
SIC2 × County × Year FE
Observations
R2
Market leverage × (Small – Large)
(t-statistic)

(1)
Full - Movers
0.050***
(2.96)
Yes
Yes
Yes
Yes
Yes
9,292,000
70.7%
-

(2)

(3)

Firm size
Small
Large
-0.113
0.063***
(-1.27)
(3.01)
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
2,336,000
6,956,000
91.5%
76.7%
0.175*
(1.87)

41

(4)
(5)
Labor market size
Small
Large
0.013
0.061***
(0.51)
(3.59)
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
2,324,000
6,969,000
84.6%
76.0%
0.049*
(1.92)

Panel B: 30% Representative Sample of All Workers (Including Movers and Non-movers, and
Including Firms with Ratings Between AA+ and B- or Unrated)
Dep. Var. = Log (earnings)
Sample:
Market leverage

(1)

(2)

All workers
0.075**
-0.006
(2.47)
(-0.33)
Yes
Yes
Yes
Yes
Yes
Yes
No
Yes
Yes
Yes
20,910,000
20,910,000
92.4%
92.5%

Worker-level controls
Firm-level controls
Worker FE
Firm FE
SIC2 × County × Year FE
Observations
R2

42

Online Appendix Table A8: Effect of Corporate Bankruptcy on Employee Earnings Conditional Analysis on Worker Mobility, Labor Market Characteristics, and Bankruptcy
Liquidation Outcomes, with UI Benefits Adjustments
This table presents the present value of earnings losses conditional on worker mobility, labor market characteristics,
and bankruptcy liquidation outcomes after estimated unemployment insurance (UI) benefits are adjusted for. “PV
(Earnings losses)” is the present value of earnings losses from years t to t+6 as a percent of the pre-bankruptcy annual
earnings in t-4. Panel A reports the present values of earnings losses conditional on worker mobility. We define a
worker as a stayer if she stays in the same firm, industry, or county through year t+3 as her t-1 firm, industry, or county.
This panel stratifies the treated (i.e., bankrupt) firm employees based on whether they stay with their firms, industries,
or counties, but not the matched control firm employees. Thus, the panel decomposes earnings losses of the treated
group workers relative to the average worker in the control group. Panel B reports the present values of earnings losses
conditional on firm size (as measured by book assets from Compustat), local labor market size (total employment in
a given two-digit SIC industry and county cell), and bankruptcy outcomes (liquidation vs. non-liquidation). We split
the sample at the 75th percentile of the firm size and labor market size distributions. This panel stratifies both the
treated and control firm employees. t-statistics for the (difference in) present value significantly differing from zero
are in parentheses (standard errors are computed using the delta method). Statistical significance at the 10%, 5%, and
1% levels is indicated by *, **, and ***.

Panel A: Present Value of Earnings Losses Conditional on Worker Mobility
(1)
(2)
(3)
(4)
(5)
Leavers:
Leavers:
Leavers:
Firm
Group:
Firm leavers same industry diff. industry same industry
stayers
same county
same county
diff. county
PV (Earnings
19.4%
-114.4%***
79.4%
26.2%
9.3%
losses)
(0.32)
(-2.90)
(1.42)
(0.53)
(0.18)
Controls
Yes
Yes
Yes
Yes
Yes
Worker FE
Yes
Yes
Yes
Yes
Yes
SIC2 × County
Yes
Yes
Yes
Yes
Yes
× Year FE

(6)
Leavers:
diff. industry
diff. county
-205.0%***
(-5.71)
Yes
Yes
Yes

Panel B: Present Value of Earnings Losses Conditional on Firm Size, Local Market Size and
Bankruptcy Outcomes
(1)
(2)
(3)
(4)
(5)
(6)
Bankruptcy liquidation
Conditioning variable:
Firm size
Labor market size
outcome
Sample:
Large
Small
Large
Small
Liquidation Non-liquidation
PV (Earnings
-46.3%
-97.7**
-20.9% -113.6%**
-95.7%*
-82.6%*
losses)
(-0.76)
(-2.30)
(-0.48)
(-2.39)
(-1.72)
(-1.95)
Diff. between two groups
51.4%
92.7%**
-13.0%
(t-statistic)
(0.96)
(2.21)
(-0.31)
Controls
Yes
Yes
Yes
Yes
Yes
Yes
Worker FE
Yes
Yes
Yes
Yes
Yes
Yes
SIC2 × County × Year FE
Yes
Yes
Yes
Yes
Yes
Yes

43