Labor Unions and Occupational Safety∗
Ling Li† Shawn Rohlin ‡ Perry Singleton §
January 12, 2019
Abstract
This study examines the causal effect of unionization on occupational safety. The em-pirical strategy exploits union election outcomes, using establishments in which elec-tions narrowly failed as a comparison group for establishments in which elections thatnarrowly passed. Data on elections come from the National Labor Relations Board,and data on occupational safety come from the Occupational Safety and Health Ad-ministration. Using accident case rates as a measure of occupational safety, the analysisfinds that unionization had no detectable effect on accident case rates at the mean, butincreased the percent of establishments with a case rate of zero by approximately 2.5percentage points. The extensive-margin effect is evident among both manufacturingand non-manufacturing establishments.
Keywords: unions, occupational safety, OSHAJEL Codes: J28, J51, J81
Corresponding Author: Perry Singleton; Syracuse University; 426 Eggers Hall;Syracuse, NY 13244; [email protected]
∗For helpful comments and suggestions, the authors would like to thank Gary Engelhardt, BrighamFrandsen, Barry Hirsch, Hugo Jales, Jeffrey Kubik, and conference participants at the Annual Meetingof the Society of Labor Economists. The authors would also like to thank Jeanette Walters-Marquez forproviding data from the Federal Mediation and Conciliation Service.†University of Wisconsin - Parkside, Department of Economics‡Kent State University, Department of Economics, and the Center for Entrepreneurship and Business
Innovation§Syracuse University, Department of Economics, and the Center for Policy Research
1 Introduction
Workers form labor unions to bargain over wages, employment, and working condi-
tions. While most research focuses on the determination of wages and employment (Farber,
1986), less research focuses on working conditions. To address this limitation, this study
examines the effect of unions on occupational safety. As Morantz (2009) notes, unions en-
gage in numerous safety-enhancing activities, including pressuring employers to maintain
safe workplaces, educating workers about workplace hazards, and developing safety-related
innovations through economies of scale. The effect on occupational safety has direct im-
plications for worker welfare and the efficiency of labor unions. The effect is also relevant
to research on unions and wages (Branchflower and Bryson, 2004; DiNardo and Lee, 2004;
Frandsen, 2014; Freeman and Medoff, 1984), since occupational safety may affect wages
through compensating differentials (Kniesner and Leeth, 2014; Rosen, 1974).
This study attempts to identify the causal effect of unionization on occupational
safety. Following DiNardo and Lee (2004), Frandsen (2014) and others, the empirical strat-
egy exploits the timing and outcome of union elections. Specifically, using the regression
discontinuity model, establishments in which elections narrowly failed are used as a compar-
ison group for establishments in which elections narrowly passed. Data on union elections
come from the U.S. National Labor Relations Board (NLRB). The analysis is limited to elec-
tions in 1991 to 2010. The data on occupational safety come from the Occupational Safety
and Health Administration (OSHA), specifically the OSHA Data Initiative (ODI). These
data report the total case rate (TCR) at the establishment level per 100 full-time equivalent
annually. The TCR includes cases involving death, days away from work, job restrictions,
job transfers, and medical attention beyond first aid. To examine union activity following an
election, the data are matched to “notices of bargaining” filed with the Federal Mediation
and Conciliation Service (FMCS).
According to the empirical analysis, unionization had no detectable effect on the
TCR at the mean, but did increase the percent of establishments with a case rate of zero by
1
approximately 2.5 percentage points. The extensive-margin effect appears to be driven by es-
tablishments that would have been at the low-end of the case-rate distribution in the absence
of unionization. Additionally, the extensive-margin effect is evident immediately after the
union election, suggesting that the effect is due to within-firm changes in occupational safety,
rather than the impact of unionization on establishment survival differentially by occupa-
tional safety. By industry, the extensive-margin effect is evident among both manufacturing
and non-manufacturing establishments.
The study contributes to an existing literature on unionization and occupational
safety. Despite the numerous safety-enhancing activities of unions, most empirical studies
find that unionization is associated with greater accidents and injuries (Donado, 2015). One
possible explanation is selection, whereby more dangerous establishments are more likely
to unionize. A few studies find that unions improve occupational safety, but these findings
pertain to specific eras and industries (Boal, 2009; Fairris, 1995; Morantz, 2013).
This study differs from related studies on two important dimensions. First, this
study exploits union elections with narrow outcomes for identification. The advantage of
this strategy is that it potentially addresses selection into unionization. A disadvantage,
however, is that the findings pertain only to newly unionized establishments with narrow
election victories. For this reason, the results form this study are not generalizable to other
contexts or directly comparable to related studies. Second, the study is the first to utilize
ODI data to examine the effect of unionization on occupational safety. This allows for a
more contemporary analysis of multiple industries, with largest shares of establishments in
manufacturing and health services.
2 Background
Workers form labor unions to create or capture monopoly rents (Farber, 1986). A
single union represents workers across multiple firms and establishments, forming national
2
and increasingly international coalitions. At the establishment level, union officials represent
workers in contract negotiations. Labor contracts generally specify compensation, but may
also specify employment levels and working conditions.
In the US, workers typically form unions through elections.1 Elections are facilitated
by the National Labor Relations Board (NLRB), established in 1935 to enforce collective
bargaining laws. To hold an election, organizers must first demonstrate at least 30 percent
support for a union election among eligible workers. This is achieved by petitions or autho-
rization cards. If successful, the NLRB determines the size and scope of the bargaining unit
and the time and location of the election. The election is conducted by secret ballot, and
a successful election requires a simple majority. If an election is successful, employers must
bargain “in good faith” with the union during contract negotiations.
A framework of union bargaining power is developed by DiNardo and Lee (2004).
In their framework, bargaining power is a function of the share of workers who favor union-
ization. In a baseline case, where union elections are permitted, but none occur, bargaining
power increases monotonically with the vote share. If an election occurs, bargaining power
increases independent of the election outcome. This is referred to as the indirect effect of
unionization. If the election is successful, bargaining power increases further. This is re-
ferred to as the direct effect of unionization. As DiNardo and Lee (2004) note, because a
successful election requires a simple majority, bargaining power increases discontinuously at
the 50-percent vote share.
In an analysis of union behavior, Farber (1986) considers two types of bargaining
structures. In the first structure, unions bargain over wages only, leaving employers to
determine employment levels and other conditions of employment. In this case, unions
negotiate along the employer’s demand curve, increasing wages and decreasing employment,
constrained by a non-negative profit condition. However, the resulting labor contract is
generally inefficient, as one party can be made better off without negatively impacting the
1An employer may independently recognize a labor union, forgoing a union election.
3
other. In the second structure, unions bargain over both wages and employment. In this
case, unions negotiate along the employer’s isoprofit curve, decreasing wages and increasing
employment relative to the demand curve. In this case, the resulting labor contract may be
efficient, but the net effect on wages and employment is ambiguous.
While most economic research on unionization has focused on wages and employ-
ment, unionization could impact other aspects of employment such as occupational safety.
Indeed, unions engage in numerous safety-enhancing activities, including pressuring employ-
ers to maintain safe workplaces, educating workers about workplace hazards, developing
safety-related innovations through economies of scale, and influencing the stringency of reg-
ulatory oversight (Morantz, 2009). A more recent study finds that unions increase OSHA
inspections, violations, and citations, which may be an important mechanism for safety
improvements (Sojourner and Yang, 2015).
Despite the safety-enhancing activities of unions, the causal effect of unions on
occupational safety is ambiguous. The conceptual framework is analogous to the framework
for wages and employment. If unions contract on wages only, employers may respond to
higher wages by decreasing investment in occupational safety. However, if unions contract
on both wages and occupational safety, occupational safety may improve. The ability to
contract on occupational safety depends, in part, on whether safety can be monitored.
The effect of unions on occupational safety is also affected by measurement and
reporting. For example, OSHA requires many employers to log workplace accidents that
require job restrictions, job transfers, or medical attention beyond first aid. One issue is
that, with unionization, employers may be more likely to accommodate workers following an
accident, thereby increasing the accident case rate. Another issue is that, with unionization,
employers may be more likely to report accidents, holding incidence constant. This may be
due to an employer’s tendency to underreport, a union’s tendency to overreport, or both.
To complicate matters further, the causal effect of unionization on occupational
safety is difficult to identify empirically. One issue is that other factors may affect both
4
occupational safety and unionization. For example, management quality may improve oc-
cupational safety and establishment profitability, with the latter increasing the likelihood of
unionization. This would generate a positive correlation between unionization and safety.
Another issue is that poor working conditions may precipitate unionization. This would
generate a negative correlation between unionization and safety. Thus, in observational
data, the correlation between unionization and occupational safety is likely biased relative
to unionization’s true causal effect.
In the empirical literature, most studies find that unionization is associated with
greater accidents and injuries (Donado, 2015). A major concern is selection, whereby es-
tablishments with greater accidents and injuries are more likely to unionize (Hills, 1985). A
few studies find positive effects of unions on occupational safety, but these findings pertain
to specific eras and industries. For example, Boal (2009) examines turn-of-the-century coal
mining, Fairris (1995) examines company unions in the 1920s, and Morantz (2013) focuses
on mining-related injuries and fatalities in the 1970s and 1980s.
3 Empirical Strategy
For identification, this study pursues an empirical strategy of DiNardo and Lee
(2004) and Frandsen (2014) and others that focuses on close union elections. In the potential
outcomes framework (Rubin, 1974; Holland, 1986), each establishment has two potential
outcomes with respect to union status: an establishment is either unionized, indicated by
Wi = 1, or not unionized, indicated by Wi = 0. The outcome with respect to union status
is denoted Yi(Wi); the causal effect among an individual establishment is denoted Yi(1) −
Yi(0); and the causal effect of unionization among all establishments is denoted E[Y (1) −
Y (0)]. Using the regression discontinuity design model (Hahn et al., 2001; Imbens and
Lemieux, 2008; Lee and Lemieux, 2010), the empirical strategy identifies the causal effect of
unionization among establishments with a vote share in favor of unionization of 50 percent,
5
the cutoff for a successful union election. This is denoted E[Y (1)−Y (0)|X = 0], where X is
the vote share relative to the 50-percent cutoff.2 The key identification assumption is that
the conditional expectation functions E[Y (1)|X] and E[Y (0)|X] are smooth at X = 0. If
so, the causal effect is given by the estimand from the sharp regression discontinuity model:
limx↓0
E[Y (1)|X]− limx↑0
E[Y (0)|X]. (1)
The left and right terms correspond to above and below the cutoff, respectively.
The causal effect can be examined both graphically and through local polynomial
regression. The regression model has the following form:
Yi = α + βWi + F (Xi) +WiG(Xi) + γZi + εi (2)
The variable Yi is a measure of occupational safety following the union election, Wi is an
indicator of union status, Xi is the vote share relative to the cutoff, and Zi is a vector of
establishment characteristics. F (.) and G(.) are polynomial functions of the vote share. By
interacting G(Xi) with Wi, the model allows for separate conditional expectation functions
above and below the cutoff. The coefficient of interest is β, which measures the discontinuity
of occupational safety at the cutoff and thus unionization’s causal effect. Because the effect
is identified locally, estimation utilizes observations only within a symmetric bandwidth
around the cutoff. The empirical analysis considers both first-order polynomials with a
narrow bandwidth and second-order polynomials with a wider bandwidth. The error term
εi is robust to heteroskedasticity.3
The effect of unionization on occupational safety may differ across the case-rate
2To impose symmetry in the vote share distribution regardless of the number of vote cast, an amountequal to 0.5 divided by the number of votes cast is subtracted from the vote share if the number of votescast is even (DiNardo and Lee, 2004).
3Discussed below, some establishments have multiple observations of occupational safety in different cal-endar years. In the empirical anlaysis, clustering the error term at the establishment level generally increasesthe standard errors, but this does not change the substantive conclusions or the statistical significance ofthe results.
6
distribution. On one hand, unions may focus their safety-enhancing efforts on high case-rate
establishments, affecting only the right tail of the case-rate distribution. On the other hand,
the right tail of the case-rate distribution may reflect idiosyncratic shocks that are unaffected
by union efforts. To estimate distributional effects, the outcome variable is replaced with an
indicator function 1(Yi ≤ y), and βy measures the discontinuity of the conditional cumulative
density function evaluated at y (Frandsen et al., 2012).
A threat to identification is non-random sorting at the cutoff. This occurs when
the vote share is manipulated at the margin of victory to alter the election outcome. Non-
random sorting would generally create discontinuities in the density of the vote share X, the
conditional distribution of establishment characteristics Z, and ultimately the conditional
distribution of the outcome of interest Y . Thus, to provide empirical support for the identi-
fication assumptions, we test for discontinuities in the density of the vote share according to
McCrary (2008) and for discontinuities in establishment characteristics using equation (2),
with Y replaced with Z.
Because the conditional distribution of covariates Zi is assumed continuous at the
cutoff, including them in equation (2) does not affect the identification strategy. Their
inclusion, however, may reduce small sample bias and improve the precision of the estimates
(Imbens and Lemieux, 2008). Given the available data, covariates include calendar year fixed
effects, industy by calendar year fixed effects, state by year fixed effects, and the number of
valid votes cast.
7
4 Data and Sample
4.1 Union Elections: National Labor Relations Board
The NLRB data on union elections come from two sources. The first is a database
compiled by the AFL-CIO, which contains elections held from 1965 to 1998.4 The second
is an online data repository, www.data.gov, which contains annual NLRB files from 1999 to
2010.5 Data from both sources include the establishment name, address, and industry, as
well as the number of eligible voters, valid votes cast, and votes for and against unionization.
Combined, the data contain 45,582 elections from 1991 to 2010.6 These years were
chosen to coincide with the years of data on occupational safety described below. The first
data source contains 21,917 elections from 1991 to 1998, and the second contains 23,665
elections from 1999 to 2010.7 The annual number of elections decreased over time, from
2,855 in 1991 to 1,644 in 2010.8 To ensure uncertainty in the election outcome, the data
are restricted to elections with at least 20 valid votes, leaving 24,758 elections from 1991 to
2010. This restriction is similarly imposed in related studies, including DiNardo and Lee
(2004), Lee and Mas (2012), Frandsen (2014), and Sojourner et al. (2015).
Table 1 provides summary statistics of the elections. The average number of votes
cast is 97.27, the average vote share in favor of unionization is 50.89 percent, and the share of
successful elections is 46.77 percent. The greatest share of the elections is in manufacturing
(28.94 percent), followed by health services (19.63 percent), transportation (16.86 percent),
4The AFL-CIO is the American Federation of Labor and the Congress of Industrial Organizations. Thedata are available to download from John-Paul Ferguson at https://github.com/jpfergongithub/nlrb oldrcases.
5The data files are labeled by calendar year, but the file name does not necessarily correspond with tallyyear of the elections within the file. After pooling the files, observations were deleted if they appeared to beduplicates or were petitions that were withdrawn or dismissed cases. For more details of the NLRB files, seethe Appendix.
6There were 455 elections that were omitted due to a missing or invalid vote share in favor of unionization.7From 1991 to 1998, the year is based on the date of the election; from 1999 to 2010, the year is bsed on
the date of the vote tally.8The number of elections is lower in 1999 compared to 1998 and 2000 - 1,686 compared to 2,761 and
2,867, respectively - suggesting some missing data in that year.
8
and construction (4.88 percent). A greater share of elections occurred in the Northeast and
Midwest, compared to the South and West.
4.2 Union Contracts: Federal Mediation and Conciliation Service
The empirical strategy assumes that bargaining power of workers increases discon-
tinuously at the 50-percent cutoff following the union election. To support this assumption,
the election data from 1999 to 2010 are matched to “notices of bargaining” data in years
1997 to 2016 from the Federal Mediation and Conciliation Service (FMCS).9 A notice is
required to initiate, terminate, or modify a labor contract and, as such, is an indicator of
union activity. The NLRB elections are matched to FMCS records by establishment name
and address.
4.3 Occupational Safety: OSHA Data Initiative
Data on occupational safety come from the ODI. The ODI was part of OSHA’s Site
Specific Targeting (SST) plan, designed to better target more dangerous establishments for
a workplace inspection. The ODI first collected a sample of accident case rates directly from
employers at the establishment level. The data were collected in annual cycles, spanning
calendar years 1996 to 2011. The sample was derived from a registry of US businesses
compiled by Dun & Bradstreet. While the sampling frame changed from cycle to cycle,
the ODI generally excluded the construction industry and smaller establishments.10 Once
collected, the data were used by the SST plan to target high case-rate establishments for an
inspection.11
The data on occupational safety come specifically from OSHA’s Form 300. This
9DiNardo and Lee (2004) similarly match union election data to the FMCS data.10In 1996, the ODI excluded establishments with fewer than 40 employees. From 1997, this threshold was
increased to 60 employees.11Using the ODI data, Li and Singleton (2018) exploit the SST plan to identify the effect of workplace
inspections on worker safety.
9
form is provided by OSHA to employers to log workplace accidents and injuries. In general,
employers with 10 or more full-time employees are required to complete the form. Cases are
logged separately involving death, days away from work, job restrictions or transfers, and
medical attention beyond first aid. Based on these logs, the ODI calculated accident case
rates per 100 full-time equivalent workers annually. The total case rate (TCR) includes all
four cases. A second rate includes only cases involving days away from work, job restrictions,
and job transfers (DART). The NLRB elections are matched to each year of the ODI based
on the establishment name and address.
Of the 24,758 union elections, 6,976 have at least one match to the ODI across all
the available years of data. Due to the years of the election and ODI data, elections closer to
1991 and 2010 were less likely to match to the ODI than elections in the intervening years.
Table 1 provides summary statistics of elections with and without a match. The number
of valid votes is greater among elections with a match, which is consistent with the ODI
excluding smaller establishments. Elections with a match are also less likely to have passed:
40.25 percent versus 49.37 percent. Regarding industry and geography, elections with a
match are more likely to be in manufacturing and health services, compared to construction
and transportation, and more likely to be in the Midwest, compared to the Northeast, South,
and West.
A single election may match to multiple ODI records in different years. Among
the 6,976 establishments with at least one match to the ODI, there are 19,318 matches from
five calendar years before the election to five calendar years after. During this period, 17.06
percent of elections have no matches, 24.68 percent have one match, 15.05 percent have two
matches, and 43.21 percent have three to eleven matches.
Figure 1 illustrates the ODI match rate each year relative to the year of the election.
The match rates are calculated using only calendar years for which ODI data are available.12
As shown, the match rate is highest in the year of the election, when the establishment is
12For example, elections tallied in 1999 were not used to calculate the ODI match rates in periods -4 and-5, which correspond to calendar years 1995 and 1994, respectively.
10
known to exist. In that year, the match rate is 11.73 percent. The match rate gradually
declines with years before and after the election, which is consistent with establishment
formation and dissolution, respectively. The match rate also reflects the ODI sampling size
and frame, which changed from cycle to cycle.13
5 Results
5.1 Vote Share
A potential threat to the identification strategy is non-random sorting at the cutoff
for a successful union election. Evidence of non-random sorting includes bunching of elections
just above or below the cutoff. To check for bunching graphically, Figure 2 plots the vote
share density relative to the cutoff, computed across 20 non-overlapping bins of 5 percentage
points each. The figure indeed suggests bunching: excluding the far-right bin, the density
increases from the right towards the cutoff, but increases only slightly in the bin just above
the cutoff, suggesting too few narrow election victories.14 The McCrary (2008) test is used
to test for bunching at the cutoff, where the null hypothesis is continuity. Using a uniform
kernel, the test fails to reject the null hypothesis at the 95 percent confidence level (p-
value=0.0696), though the test statistic is only marginally insignificant.
Manipulating the vote share may become increasingly more difficult as the number
of voters increases. While Figure 2 illustrates the vote share density among elections with at
least 20 votes, Figure 3 illustrates the vote share densities using greater vote thresholds. As
shown, the lower density just above the cutoff is no longer apparent among elections with
13Because the sampling frame changed from cycle to cycle, matched ODI observations are not directlycomparable across calendar years or analysis periods (Figure 1). This prevents event-study analysis, whichcompares changes in the mean case rate before and after the union election.
14Frandsen (2014) also finds evidence of bunching, suggesting too few narrow election victories. A keydifference is that his data include earlier years, from 1980 to 2009. In our data, evidence of bunching isstronger in earlier years (e.g. 1992 to 1999) than later years (e.g. 2000 to 2010).
11
greater votes. Moreoever, the McCrary (2008) test consistently fails to reject continuity.15
Thus, a potential strategy to address non-random sorting at the cutoff is to focus on union
elections with greater votes cast.
5.2 Establishment Characteristics
Non-random sorting may also lead to discontinuities in the distribution of estab-
lishment and election characteristics. Regarding election characteristics, Figure 4 plots the
conditional mean of eligible employees and valid votes cast.16 As shown, both measures
increase and then decrease with the vote share, with no apparent discontinuity at the cutoff.
Table 2 presents discontinuity estimates using equation (2). The rows correspond to different
outcome variables: the first row is eligible employees; the second row is valid votes cast. Each
column corresponds to a discontinuity estimate from a single model: In the first column, the
model utilizes a polynomial of order one (linear) and a bandwidth of 15 percentage points; in
the second column, the model utilizes a polynomial of order is two (quadratic) and a band-
width of 25 percentage points. As shown, all the discontinuity estimates are small compared
to the mean near the cutoff, and none are statistically significant. In the first column, the
estimated discontinuity for eligible employees is 2.41, compared to a mean of 116.98, and the
estimated discontinuity for valid votes is 3.06, compared to a mean of 102.90.17
Figure 5 illustrates the percent of establishments in manufacturing and health ser-
vices. As shown, the share in manufacturing increases and then decreases with the vote
share, whereas the share in health services generally increases, except for a sharp decrease
at the highest bins. Graphically, neither measure of industry exhibits a discontinuity at the
cutoff. Additionally, the discontinuity estimates presented in the third and fourth rows of
15Using the uniform kernel, the p-values are 0.241, 0.664, 0.583, and 0.712 using a vote threshold of 30,50, 70, and 100, respectively.
16When examining discontinuities in the number of eligible employees and votes cast, one extreme outlieris omitted from the analysis. For this outlier, the number of eligible employees and votes case is 17,195 and15,471, respecitively. The next highest values are 7,000 and 4,589, respectively.
17The means are calculated among elections within plus or minus 10 percentage points from the cutoff.
12
Table 2 are small and statistically insignificant.
A final measure of establishment characteristics is whether an establishment matches
to the ODI. Stated above, a match to the ODI reflects, in part, the sampling size and frame,
which varied from cycle to cycle. Figure 6 illustrates the percent of firms that match to the
ODI. As shown, the match rate increases and then decreases with the vote share, with no
apparent discontinuity at the cutoff. Additionally, the discontinuity estimate presented in
the fifth row of Table 2 is small and statistically insignificant.
Taken together, the results suggest that establishments are comparable just above
and below the cutoff with respect to establishment and election characteristics. These find-
ings are consistent with the identification assumption that the conditional expectation func-
tions of occupational safety, as well as both observable and unobservable factors that affect
occupational safety, are smooth at the cutoff. If so, the estimated discontinuity of occupa-
tional safety at the cutoff may be interpreted as the causal effect of unionization.
5.3 Union Activity
The identification strategy assumes that bargaining power of workers increases dis-
continuously at the cutoff for a successful union election. Using the FMCS data on notices
of bargaining, Figure 7 plots the FMCS match percent by calendar year before and after
the union election, separately by the election outcome. Among establishments in which the
election passed, the match rate increases sharply in the calendar year of the election and
the year after, then returns to its pre-existing trend. Among establishments in which the
election failed, the match rate remains relatively unchanged compared to the pre-existing
trend. These results confirm that union activity increases following a successful election.
To examine union activity at the cutoff, Figure 8 plots the FMCS match percent
in periods 0 and 1 by vote share. These periods had the greatest increase in union activity
following a successful union election, according to Figure 7. As shown, union activity in-
creases discontinuously at the cutoff following a union election. The discontinuity estimates
13
presented in the sixth row of Table 2 are 34.63 percentage points using the first model and
32.37 percentage points using the second model. Both estimates are statistically significant
at the one percent level.
5.4 Occupational Safety
5.4.1 After Election
The discontinuous increase in union activity at the cutoff is used to identify the
effect of unionization on occupational safety. Using NLRB-ODI matched observations, the
empirical analysis focuses intially on the TCR after the election. To increase the sample
size, the data are pooled across periods one through five, corresponding to the first through
fifth calendar years after the election. The pooled data contain 10,835 observations.
The left panel of Figure 9 illustrates the mean TCR by vote share. As shown, the
mean TCR generally increases with the vote share, except for a decrease at the highest bins.
The estimates appear noisier towards the extremes of the vote share, where the density of
the vote share is lower. Importantly, there is no apparent discontinuity in the mean TCR at
the cutoff. Table 3 reports the discontinuity estimates using equation (2). In the first row,
the outcome variable is the TCR, and each column corresponds to an estimate from a single
model. In the first column, the polynomial order is one, and the bandwidth is 15 percentage
points. In this case, the discontinuity estimate is -0.0770. Although the estimate is negative,
which suggests that unionization improved occupational safety, it is small and statistically
insignificant. The point estimate is similar if the model includes covariates, as reported in
the second column. In the third column, the polynomial order is two (quadratic), and the
bandwidth is 25 percentage points. In this specification, the discontinuity estimate is -0.394,
which is also negative, but statistically insignificant. Again, the point estimate is similar if
the model includes covariates, as reported in the fourth column. While the results suggest
that unionization does not affect occupational safety at the mean, the standard errors do not
14
rule out a wide range of effects. For example, in the first column, the 95 percent confidence
interval is -1.25 to 1.10, compared to a mean near the cutoff of approximately 14.
Stated above, unionization may affect occupational safety at different parts of the
case-rate distribution. To examine occupational safety on the extensive margin, the right
panel of Figure 9 illustrates the percent of establishments in which the TCR is zero. As
shown, the percent with zero cases decreases and then increases with the vote share, reach-
ing a minimum near the cutoff. Importantly, the percent with zero cases increases discon-
tinuously at the cutoff, suggesting that unionization improved occupational safety on the
extensive margin. The discontinuity estimates with 1(TCR = 0) as the outcome variable
are reported in the second row of Table 3. As shown, the estimates are similar across all
models, ranging from 2.521 percentage points to 2.852 percentage points. All the estimates
are statistically significant at the five percent level, and the estimate in the first column is
significant at the one percent level.
If unionization improved occupational safety on the extensive margin, but had
no detectable effect at the mean, then the extensive-margin effects likely occurred among
establishments that would have been at the low-end of the case-rate distribution in the
absence of unionization. If the effects occurred at the high-end, instead, the effect at the
mean would be larger. To confirm this point, Figure 10 illustrates discontinuity estimates of
the conditional cumulative density function at each integer of the TCR distribution, from 1
to 25.18 The models in the left panel have a first-order polynomial and a bandwidth of 15
percentage points, and the models in the right panel have a second-order polynomial and
a bandwidth of 25 percentage points. The models in both panels include covariates. As
shown, the discontinuity estimates are positive and statistically significant from 1 to 4, but
decrease and become statistically insignificant thereafter. From 5 to 25, most of the estimates
are small, and all of the estimates are statistically insignificant. These results confirm that
the shift in the case-rate distribution towards zero occurred at the bottom of the case-rate
18A TCR at 25 corresponds to the 90th percentile of the case-rate distribution.
15
distribution.
5.4.2 Before Election
To establish causality, it is important to consider whether there is a discontinuity
in occupational safety at the cutoff before the election. Figure 11 illustrates the TCR by
vote share before the election, using NLRB-ODI matched observations in periods negative
five through negative one. As shown, the estimates before the election appear somewhat
noisier than the those after the election. This is due, in part, to fewer NLRB-ODI matched
observations before the election: 6,445 versus 10,835.19 Graphically, there is no apparent
discontinuity at the mean in the left panel or at the extensive margin in the right panel.
The third row of Table 3 presents discontinuity estimates for the TCR at the mean. While
the estimates in the first and third columns suggest a negative discontinuity at the cutoff,
only the estimate in the first column is statistically significant at the five percent level, and
neither estimate is robust to the inclusion of covariates, as shown in the second and fourth
columns. Moreover, the discontinuity estimates are sensitive to outliers in the case-rate
distribution. For example, by winsorizing the case rate at the 99th percent, the estimate
in the first column decreases from -1.801 with a standard error of 0.897 to -1.337 with a
standard error of 0.827, which is statistically insignificant. By windsorizing at the 95th
percentile, the estimates decreases further to -0.727 with a standard error of 0.701. The
fourth row of Table 3 presents discontinuity estimates of the TCR on the extensive margin.
As shown, all the estimates are negative and statistically insignificant. The estimates are
also sensitive to the model specification, appearing smaller in the model with a second order
polynomial and the wider bandwidth (columns three and four).
To examine distributional effects before the election, Figure 12 illustrates disconti-
nuity estimates of the conditional cumulative density function of the TCR, similar to Fig-
ure 10. As shown, most estimates are relatively small and statistically insignificant. Impor-
19The difference in matched observations before and after the election partially reflects that the rate offirm formation before the election is higher than the rate of firm dissolution after the election (Figure 1).
16
tantly, the discontinuity estimates at one through four, where the estimates are positive and
statistically significant after the election, are small and statistically insignificant before the
election. This suggests that the extensive margin effects are causal, having arisen only after
a union election.
5.4.3 DART
The ODI data also report the DART rate, a subset of the TCR. Specifically, the
DART includes cases involving days away from work, job restrictions, job transfers, and
excludes cases involving death or medical attention beyond first aid. Table 4 reports discon-
tinuity estimates for the DART at the mean and on the extensive margin both before and
after the election. After the election, the results for the DART are qualitatively similar to
the results for the TCR. The mean effects are mostly negative, though statistically insignif-
icant, and the extensive margin effects are positive and statistically significant (Figure 13).
Before the election, the results for the DART differ slightly from the results for the TCR.
The mean effects are positive, but small and statistically insignificant, and the extensive
margin are sizeable, ranging from -2.054 to -2.724, but are statistically insignificant. By
examining the extensive margin effects graphically, they appear to reflect sampling noise,
rather than systemic differences between establishments just above and below the cutoff
(Figure 14). Because the results for the TCR and the DART are qualitatively similar, and
because the DART is a subset of the TCR, the effect of unionization on occupational safety
predominately stems from cases involving days away from work, job restrictions, and job
transfers.
5.4.4 Firm Size
When examining the vote share distribution, there was concern for non-random
sorting at the cutoff, specifically among establishments with fewer than 30 valid votes. To
address this issue, the empirical analysis is repeated using only establishments with 30 or
17
more valid votes. The results are qualitatively similar with the higher cutoff. For example,
with a cutoff of 20, the discontinuity estimate on the extensive margin, reported in column
two of Table 3, is 2.521 percentage points, which is statistically significant at the five percent
level. With a cutoff of 30, the discontinuity estimate is 2.404, which is also statistically
significant at the five percent level. Because non-random sorting is arguably less of a concern
among larger establishments, the discontinuity estimates using the higher cutoff are more
likely to reflect the causal effect of unionization.
5.4.5 Firm Survival
To properly interpret the results, it is important to consider whether unionization
affects firm survival. While DiNardo and Lee (2004) and Freeman and Kleiner (1999) find
no effects of unionization on firm survival, Brown and Heywood (2006) and Frandsen (2014)
suggest unionization decreases firm survival, evident three years after the union election. If
so, a particular issue is whether unionization affects firm survival differentially by occupa-
tional safety. For example, unionization may cause more dangerous firms to dissolve, which
would decrease the average case rate and increase the percent of establishments with a case
rate of zero, independent of within-establishment effects of unionization on occupational
safety.
To address this issue, the analysis is repeated using only observations shortly after
the election, before the effect of unionization on firm survival is evident. When using periods
1 through 5, with a sample size of 5,499, the discontinuity estimate on the extensive margin,
reported in column two of Table 3, is 2.521 percentage points and statistically significant
at the five percent level. When using periods 1 through 2, with a sample of 2,200, the
discontinuity estimate is 2.292, but statistically insignificant. When using periods 1 through
3, with a sample of 3,327, the discontinuity estimate is 2.804 and statistically significant.
Because the extensive-margin effects are evident in the short run, the effect of unionization on
occupational safety likely reflects within-establishment changes in ocupational safety, rather
18
than differential impacts on firm survival with respect to occupational safety.
5.4.6 Multiple Elections
A single establishment may have multiple elections. For example, among the 45,582
elections from 1991 to 2010, there are 35,643 unique establishments based on an establish-
ment name, state, and city. 89.61 percent of establishments had only one election, and
8.13 percent had two elections. One concern with the preceding analysis is that multiple
elections, as well as their individual matches to the ODI, were counted as multiple observa-
tions. Another concern is that an establishment in which an election barely fails may have a
subsequent election that is successful, which would contaminate the comparison group with
treatment several years after the election.
To address this issue, the analysis is restricted to elections with no other elections
five periods before or after. The results are qualitatively similar. When using all elections,
with a sample size of 5,499, the discontinuity estimate on the extensive margin, reported
in column two of Table 3, is 2.521 percentage points and statistically significant at the five
percent level. When using the added restirction, with a sample size of 3,929, the discontinuity
estimate is 3.106 percentage points and statistically significant.
5.4.7 Effects by Industry
To examine whether the extensive margin effects vary by industry, the analysis is
repeated separately for manufacturing and non-manufacturing, which approximately halves
the sample. Figure 15 illustrates the percent of establishments in which the TCR is zero. As
shown, the percent with zero cases appears to increase discontinuously at the cutoff for both
industry categories. When using all elections, with a sample size of 5,499, the discontinuity
estimate on the extensive margin, reported in column two of Table 3, is 2.521 percentage
points and statistically significant. When using all elections in manufacturing, with a sample
size of 2,987, the discontinuity estimate is 3.222 percentage points and statistically significant
19
at the five percent level. When using all elections not in manufacturing, with a sample size
of 2,512, the discontinuity estimate is 3.826 percentage points and statistically significant.
Thus, the extensive-margin effects of unionization on occupational safety are evident in both
manufacturing and non-manufacturing establishments.
6 Conclusion
Unions engage in numerous safety-enhancing activities. To identify the effect of
unionization on occupational safety, this study exploits union election outcomes, using es-
tablishments in which elections narrowly failed as a comparison group for establishments
in which elections narrowly passed. While unionization had no detectable effect on occu-
pational safety at the mean, unionization did improve occupational safety on the extensive
margin, increasing the share of establishments with an annual accident case rate of zero by
approximately 2.5 percentage points. According to auxiliary analysis, the extensive-margin
effect appears to have occurred among establishments that would have been at the low end
of the case-rate distribution in the absence of unionization.
The results have direct implications for research on unions and wages, since occu-
pational safety may affects wages through compensating differentials (Kniesner and Leeth,
2014; Rosen, 1974). According to the analysis, the extensive-margin effects occurred among
establishments that would have had an annual TCR of four or less in the absence of union-
ization. Thus, as an upper bound among these establishments, the case rate could have
decreased by as much as four per 100 full-time equivalent workers annually. The implicit
value of statistical injury ranges from $33 thousand to $182 thousand in 2010 dollars (Viscusi
and Aldy, 2003), so a reduction of four cases is valued at $132 thousand to $728 thousand.
This translates into an increase in the hourly wage between $0.66 and $3.64, based on a
full-time work schedule of 2,000 hours per year. The median hourly wage among unionized
workers in manufacturing was $20.70 (Bureau of Labor Statistics, 2011). Thus, as an upper
20
bound among affected establishments, wages could be as much as 17.6 percent higher after
accounting for improvements in occupational safety.
21
Appendix
The study uses data from multiple sources. Data on union elections come from
the National Labor Relations Board (NLRB) and the American Federation of Labor and
the Congress of Industrial Organizations; data on union activity come from the Federal
Mediation and Conciliation Service (FMCS); and data on occupational safety come from
the Occupational Safety and Health Administration (OSHA), specifically the OSHA Data
Initiative (ODI). These data are matched at the establishment level.
The NLRB reports the establishment name, address, and industry, as well as the
number of eligible voters, valid votes cast, and votes for and against unionization. The
NLRB data are first restricted to closed cases. To match the NLRB data to the other data,
the establishment name and address were standardized. For the establishment name, all
the special characters and common words, such as company, limited, and corporation, were
deleted. If the listed formal name and the case name differed, or if the establishment is
“doing business as (DBA) under a different name, both names are retained and used for
matching. For the street address, all special characters and numbers for floor, suite, and
room were deleted. Common words, such as street, avenue, and road, were replaced with
their respective abbreviations. To standardize and clean the city name, each name was best
matched to an exhaustive list of all city names in the US, compiled by the US Census Bureau.
City names without a perfect match were checked manually for typos.
The FMCS data indicate whether an establishment filed a ”notice of bargaining,”
an indicator of union activity. The ODI reports accident case rates, measured annually
per 100 full-time equivalent workers. To match the FMCS and the ODI to the NLRB, the
establishment name and address were standardized using the same method as the NLRB.
The matching procedure utilized the establishment name, street address, city, state, and
zip code. The NLRB was matched to the FMCS and ODI in several stages. In the first
stage and most restrictive stage, the data were matched on the establishment name, street,
city, and state. In the second stage, the data were matched on the establishment name, zip
22
code, city, and state. In the third stage, the data were matched based on the first six letters
of the establishment name and address. If an FMCS or ODI record successfully matched
in one stage, the matched record was removed from matching in subsequent stages. If a
record contained multiple establishment names, the matching procedure was repeated for
each name until a successful match, if any.
23
References
Boal, W. (2009). The effect of unionism on accidents in u.s. coal mining, 1897-1929. Industrial
Relations, 48:97–120.
Branchflower, D. and Bryson, A. (2004). What effect do unions have on wages now and
would freeman and medoff be surprised? Journal of Labor Research, 25:383–414.
Brown, M. and Heywood, J. (2006). Investigating the cause of death: Industrial relations
and plant closures in australia. Industrial and Labor Relations Review, 59(4):593–612.
Bureau of Labor Statistics (2011). Union members - 2010. New Release USDL-11-0063.
DiNardo, J. and Lee, D. (2004). Economic impacts of new unionization on private sector
employers: 1984-2001. Quarterly Journal of Economics, 119:1383–1442.
Donado, A. (2015). Why do unionized workers have more nonfatal occupational injuries?
Industrial and Labor Relations Review, 68:153–183.
Fairris, D. (1995). From exit to voice in shopfloor governance: The case of company unions.
Business History Review, 69:494–529.
Farber, H. (1986). The analysis of union behavior. In Ashenfelter, O. and Layard, R., editors,
Handbook of Labor Economics, volume 2, pages 1039–1089. Elsevier.
Frandsen, B. (2014). The surprising impacts of unionization: Evidence from matched
employer-employee data. Unpublished paper.
Frandsen, B., Frolich, M., and Melly, B. (2012). Quantile treatment effects in the regression
discontinuity design. Journal of Econometrics, 168:382–395.
Freeman, R. and Kleiner, M. (1999). Do unions make enterprises insolvent? Industrial and
Labor Relations Review, 52(4):510–527.
Freeman, R. and Medoff, J. (1984). What Do Unions Do? Basic Books, New York.
24
Hahn, J., Todd, P., and der Klaauw, W. V. (2001). Identification and estimation of treatment
effects with a regression-discontinuity design. Econometrica, 69:201–209.
Hills, S. (1985). The attitudes of union and nonunion male workers towards union represen-
tation. Industrial and Labor Relations Review, 38:179–194.
Holland, P. (1986). Statistics and causal inference. Journal of the American Statistical
Association, 81:945–960.
Imbens, G. and Lemieux, T. (2008). Regression discontinuity designs: A guide to practice.
Journal of Econometrics, 142:615–635.
Kniesner, T. and Leeth, J. (2014). Regulating occupational and product risks. In Machina,
M. and Viscusi, W., editors, Handbook of Economics of Risk and Uncertainty, volume 1,
pages 493–600. Elsevier.
Lee, D. and Lemieux, T. (2010). Regression discontinuity designs in econometrics. Journal
of Economic Literature, 48:281–355.
Lee, D. and Mas, A. (2012). Long-run impacts of unions on firms: New evidence from
financial markets, 1961-1999. Quarterly Journal of Economics, 127:333–378–529.
Li, L. and Singleton, P. (2018). The effect of workplace inspection on worker safety. Unpub-
lished paper.
McCrary, J. (2008). Manipulation of the running variable in the regression discontinuity
design: A density test. Journal of Econometrics, 142:698–714.
Morantz, A. (2009). The elusive union safety effect: Toward a new empirical research agenda.
In Eaton, A., editor, Labor and Employment Relations Association, 61st Proceedings, pages
130–146.
Morantz, A. (2013). Coal mine safety: Do unions make a difference? Industrial and Labor
Relations Review, 66:87–116.
25
Rosen, S. (1974). Hedonic prices and implicit markets: Produce differentiation in pure
competition. Journal of Political Economy, 82(1):34–55.
Rubin, D. (1974). Estimating causal effects of treatments in randomized and non-randomized
studies. Journal of Educational Psychology, 66:688–701.
Sojourner, A., Town, R., Grabowski, D., Chen, M., and Frandsen, B. (2015). Greasing the
squeaky wheel: The relative productivity of osha complaint inspections. Industrial and
Labor Relations Review, 64(4):771–806.
Sojourner, A. and Yang, J. (2015). Effects of unionization on workplace-safety enforcement:
Regression-discontinuity evidence. IZA Discussion Paper 9610.
Viscusi, K. and Aldy, J. (2003). The value of a statistical life. Journal of Risk and Uncer-
tainty, 27:5–76.
26
Table 1: Summary Statistics by Match to ODI
ODIMatch
All Yes No
Eligible employees (number) 114.0 139.0 104.1(1.381) (2.696) (1.598)
Valid votes (number) 97.27 122.7 87.13(1.139) (2.252) (1.309)
Vote share 50.89 47.70 52.16(0.151) (0.261) (0.183)
Pass 46.77 40.25 49.37(0.319) (0.587) (0.378)
Construction 4.884 3.813 5.310(0.138) (0.229) (0.169)
Manufacturing 28.94 47.91 21.38(0.290) (0.598) (0.310)
Transportation 16.86 9.389 19.84(0.239) (0.349) (0.301)
Health services 19.63 21.44 18.91(0.254) (0.491) (0.296)
Other 29.68 17.45 34.56(0.292) (0.454) (0.359)
Northeast 27.08 25.73 27.61(0.284) (0.523) (0.338)
Midwest 29.77 37.28 26.78(0.292) (0.579) (0.335)
South 22.41 20.83 23.03(0.266) (0.486) (0.318)
West 20.74 16.16 22.57(0.259) (0.441) (0.316)
Elections 24,758 6,976 17,782
The table presents summary statistics of union elections. The sample is derived from union
elections contained in the NLRB, file years 1991 to 2010. The sample is restricted to elections
with at least 20 valid votes and a valid vote share in favor of unionization. The second and
third columns present summary statistics separately by whether the union election is matched
to any observations in the OSHA Data Initiative (ODI), file years 1996 to 2011. Standard
errors are in parentheses.
27
Table 2: Discontinuity Estimates of Establishment Characteristics
Outcome Variable (1) (2)Employees eligible (number) 2.412 -1.360
(6.733) (7.883)[11,149] [17,186]
Valid votes (number) 3.059 1.535(5.626) (6.600)[11,149] [17,186]
Manufacturing -0.802 -0.682(1.779) (2.100)[11,150] [17,187]
Health services 0.884 1.127(1.582) (1.865)[11,150] [17,187]
ODI match 0.507 0.0241(1.780) (2.085)[11,150] [17,187]
FMCS match in periods 0 and 1 34.63*** 32.37***(2.482) (2.940)[5,365] [8,282]
Polynomial 1 2Bandwidth 0.15 0.25
The table presents discontinuity estimates of establishment characteristics. The sample is
derived from union elections contained in the NLRB, file years 1991 to 2010. The sample
is restricted to elections with at least 20 valid votes and a valid vote share in favor of
unionization. One outlier is dropped when the outcome variable is employees eligible and
valid votes. Robust standard errors are in parentheses, and sample sizes are in brackets.
Estimates are in percentage points unless otherwise noted. ∗∗∗, ∗∗, and ∗ indicate significance
at the one, five, and ten percent levels, respectively.
28
Table 3: Discontinuity Estimates of TCR
Outcome Variable Periods (1) (2) (3) (4)TCR 1 through 5 -0.0770 -0.0939 -0.394 -0.344
(0.599) (0.608) (0.697) (0.689)[5,499] [5,499] [8,303] [8,303]
TCR=0 1 through 5 2.578*** 2.521** 2.587** 2.852**(0.991) (0.991) (1.196) (1.176)[5,499] [5,499] [8,303] [8,303]
TCR -5 through -1 -1.801** -1.215 -1.884* -1.018(0.897) (0.967) (1.067) (1.079)[3,025] [3,025] [4,617] [4,617]
TCR=0 -5 through -1 -1.915 -1.878 -0.906 -0.894(1.317) (1.516) (1.640) (1.758)[3,025] [3,025] [4,617] [4,617]
Polynomial 1 1 2 2Bandwidth 0.15 0.15 0.25 0.25Covariates No Yes No Yes
The table presents discontinuity estimates of the TCR. The sample is derived from union
elections contained in the NLRB, file years 1991 to 2010. The sample is restricted to elections
with at least 20 valid votes and a valid vote share in favor of unionization. The total case rate
(TCR) includes cases involving death, days away from work, job restrictions, job transfers,
and medical attention beyond first aid. The case rate is measured per 100 full-time equivalent
workers annually. Robust standard errors are in parentheses. The estimates for 1(TCR = 0)
are reported in percentage points. ∗∗∗, ∗∗, and ∗ indicate significance at the one, five, and
ten percent levels, respectively.
29
Table 4: Discontinuity Estimates of DART
Outcome Variable Periods (1) (2) (3) (4)DART 1 through 5 0.0668 -0.164 -0.131 -0.450
(0.395) (0.406) (0.461) (0.464)[5,499] [5,499] [8,303] [8,303]
DART=0 1 through 5 3.995*** 3.440*** 4.685*** 4.196***(1.318) (1.330) (1.580) (1.570)[5,499] [5,499] [8,303] [8,303]
DART -5 through -1 0.184 0.322 0.424 0.724(0.497) (0.542) (0.579) (0.590)[3,025] [3,025] [4,617] [4,617]
DART=0 -5 through -1 -2.724 -2.446 -2.419 -2.054(1.767) (1.974) (2.183) (2.339)[3,025] [3,025] [4,617] [4,617]
Polynomial 1 1 2 2Bandwidth 0.15 0.15 0.25 0.25Covariates No Yes No Yes
The table presents discontinuity estimates of the DART. The sample is derived from union
elections contained in the NLRB, file years 1991 to 2010. The sample is restricted to elections
with at least 20 valid votes and a valid vote share in favor of unionization. The DART
includes cases involving days away from work, job restrictions, and job transfers. The case
rate is measured per 100 full-time equivalent workers annually. Robust standard errors are
in parentheses. The estimates for 1(DART = 0) are reported in percentage points. ∗∗∗, ∗∗,
and ∗ indicate significance at the one, five, and ten percent levels, respectively.
30
Figure 1: ODI Match Rate by Period
05
1015
20M
atch
Per
cent
-5 -4 -3 -2 -1 0 1 2 3 4 5Period
The figure illustrates the match rate of union elections contained in the NRLB to occupa-
tional safety data contained in the ODI. The sample is derived from union elections contained
in the NLRB, file years 1991 to 2010. The sample is restricted to elections with at least 20
valid votes and a valid vote share in favor of unionization. The periods correspond to calen-
dar years relative to the calendar year of the election. The match rates are calculated using
only calendar years for which ODI data are available. For example, elections tallied in 1999
were not used to calculate the ODI match rates in periods -4 and -5, which correspond to
calendar years 1995 and 1994, respectively.
31
Figure 2: Distribution of Vote Share
0.5
11.
52
2.5
3D
ensi
ty
-.5 -.25 0 .25 .5Vote Share
Full Sample
The figure illustrates the vote share distribution in favor of unionization. The sample is
derived from union elections contained in the NLRB, file years 1991 to 2010. The sample
is restricted to elections with at least 20 valid votes and a valid vote share in favor of
unionization.
32
Figure 3: Distribution of Vote Share
0.5
11.
52
2.5
3D
ensi
ty
-.5 -.25 0 .25 .5Vote Share
Votes>=30
0.5
11.
52
2.5
3D
ensi
ty
-.5 -.25 0 .25 .5Vote Share
Votes>=50
0.5
11.
52
2.5
3D
ensi
ty
-.5 -.25 0 .25 .5Vote Share
Votes>=70
0.5
11.
52
2.5
3D
ensi
ty
-.5 -.25 0 .25 .5Vote Share
Votes>=100
The figure illustrates the vote share distribution in favor of unionization. The sample is
derived from union elections contained in the NLRB, file years 1991 to 2010. The sample
is restricted to elections with at least 20 valid votes and a valid vote share in favor of
unionization. In each panel, the sample is restricted to elections with a minimum number of
valid votes cast, as indicated.
33
Figure 4: Employees and Votes by Vote Share
050
100
150
200
250
Num
ber
-.5 0 .5Vote Share
Eligible Employees
050
100
150
200
250
Num
ber
-.5 0 .5Vote Share
Valid Votes
The figure illustrates the average number of eligible voters and valid votes cast for 20 non-
overlapping bins of 5 percentage points each. The sample is derived from union elections
contained in the NLRB, file years 1991 to 2010. The sample is restricted to elections with
at least 20 valid votes and a valid vote share in favor of unionization.
34
Figure 5: Industry by Vote Share
010
2030
4050
6070
8090
100
Perc
ent
-.5 0 .5Vote Share
Manufacturing
010
2030
4050
6070
8090
100
Perc
ent
-.5 0 .5Vote Share
Health Services
The figure illustrates the share of establishment in manufacturing and health services for 20
non-overlapping bins of 5 percentage points each. The sample is derived from union elections
contained in the NLRB, file years 1991 to 2010. The sample is restricted to elections with
at least 20 valid votes and a valid vote share in favor of unionization.
35
Figure 6: ODI Match by Vote Share
010
2030
4050
6070
8090
100
Perc
ent
-.5 0 .5Vote Share
ODI Match
The figure illustrates the match rate of union elections to the ODI for 20 non-overlapping bins
of 5 percentage points each. The ODI contains data on occupational safety and is compiled
by the Occupational Safety and Health Administration. The sample is derived from union
elections contained in the NLRB, file years 1991 to 2010. The sample is restricted to elections
with at least 20 valid votes and a valid vote share in favor of unionization.
36
Figure 7: FMCS Match by Period
010
2030
4050
Mat
ch P
erce
nt
-5 -4 -3 -2 -1 0 1 2 3 4 5Period
Pass Fail
The figure illustrates the match rate of union elections contained in the NRLB to notices of
bargaining filed with the FMCS. The sample is derived from union elections contained in the
NLRB, file years 1999 to 2010. The sample is restricted to elections with at least 20 valid
votes and a valid vote share in favor of unionization. The periods correspond to calendar
years relative to the calendar year of the election.
37
Figure 8: FMCS Match by Vote Share
010
2030
4050
6070
8090
100
Mat
ch P
erce
nt
-.5 0 .5Vote Share
The figure illustrates the match rate of union elections contained in the NRLB to notices of
bargining filed with the FMCS in periods zero or one across bins of five percentage points.
The sample is derived from union elections contained in the NLRB, file years 1999 to 2010.
The sample is restricted to elections with at least 20 valid votes and a valid vote share in
favor of unionization.
38
Figure 9: TCR by Vote Share, Periods 1 through 5
010
2030
TCR
/100
full-
time
empl
oyee
s an
nual
ly
-.5 0 .5Vote Share
TCR
010
2030
Perc
ent A
nnua
l TC
R=0
-.5 0 .5Vote Share
TCR=0
The figure illustrates the TCR in periods 1 to 5 for 20 non-overlapping bins of 5 percentage
points each. The sample is derived from union elections contained in the NLRB, file years
1991 to 2010. The sample is restricted to elections with at least 20 valid votes and a valid
vote share in favor of unionization.
39
Figure 10: Discontinuity Estimates of the Cumulative Density Function ofthe TCR, Periods 1 through 5
-30
-20
-10
010
2030
Dis
cont
inui
ty E
stim
ate
0 5 10 15 20 25Outcome: I(TCR
Figure 11: TCR by Vote Share, Periods -5 through -1
010
2030
TCR
/100
full-
time
empl
oyee
s an
nual
ly
-.5 0 .5Vote Share
TCR
010
2030
Perc
ent A
nnua
l TC
R=0
-.5 0 .5Vote Share
TCR=0
The figure illustrates the mean TCR in periods -5 to -1 for 20 non-overlapping bins of 5
percentage points each. The sample is derived from union elections contained in the NLRB,
file years 1991 to 2010. The sample is restricted to elections with at least 20 valid votes and
a valid vote share in favor of unionization.
41
Figure 12: Discontinuity Estimates of the Cumulative Density Function ofthe TCR, Periods -5 to -1
-30
-20
-10
010
2030
Dis
cont
inui
ty E
stim
ate
0 5 10 15 20 25Outcome: I(TCR
Figure 13: DART by Vote Share, Periods 1 through 5
010
2030
DAR
T/10
0 fu
ll-tim
e em
ploy
ees
annu
ally
-.5 0 .5Vote Share
DART
010
2030
Perc
ent A
nnua
l DAR
T=0
-.5 0 .5Vote Share
DART=0
The figure illustrates the DART in periods 1 to 5 for 20 non-overlapping bins of 5 percentage
points each. The sample is derived from union elections contained in the NLRB, file years
1991 to 2010. The sample is restricted to elections with at least 20 valid votes and a valid
vote share in favor of unionization.
43
Figure 14: DART by Vote Share, Periods -5 through -1
010
2030
DAR
T/10
0 fu
ll-tim
e em
ploy
ees
annu
ally
-.5 0 .5Vote Share
DART
010
2030
Perc
ent A
nnua
l DAR
T=0
-.5 0 .5Vote Share
DART=0
The figure illustrates the DART in periods -5 to -1 for 20 non-overlapping bins of 5 percentage
points each. The sample is derived from union elections contained in the NLRB, file years
1991 to 2010. The sample is restricted to elections with at least 20 valid votes and a valid
vote share in favor of unionization.
44
Figure 15: Discontinuity Estimates of the Cumulative Density Function ofthe TCR, Periods 1 to 5
010
2030
Perc
ent A
nnua
l TC
R=0
-.5 0 .5Vote Share
Manufacturing
010
2030
Perc
ent A
nnua
l TC
R=0
-.5 0 .5Vote Share
Not Manufacturing
The figure illustrates the TCR in periods 1 to 5 for 20 non-overlapping bins of 5 percentage
points each. The sample is derived from union elections contained in the NLRB, file years
1991 to 2010. The sample is restricted to elections with at least 20 valid votes and a valid
vote share in favor of unionization.
45