LONG-RUN IMPACTS OF UNIONS ON FIRMS:
NEW EVIDENCE FROM FINANCIAL MARKETS, 1961-1999∗
David S. Lee
Princeton University and NBER
Alexandre Mas
Princeton University and NBER
May 2011
Abstract
We estimate the effect of new private-sector unionization on publicly-traded ﬁrms’ equity value in
the U.S. over the 1961-1999 period using a newly assembled sample of National Labor Relations Board
(NLRB) representation elections matched to stock market data. Event-study estimates show an average
union effect on the equity value of the ﬁrm equivalent to $40,500 per unionized worker, an effect that
takes 15 to 18 months after unionization to fully materialize, and one that could not be detected by a shortrun
event study. At the same time, point estimates from a regression-discontinuity design – comparing
the stock market impact of close union election wins to close losses – are considerably smaller and close
to zero. We ﬁnd a negative relationship between the cumulative abnormal returns and the vote share in
support of the union, allowing us to reconcile these seemingly contradictory ﬁndings. JEL Codes: J01,
J08, J5, J51
∗We thank Jonathan Berk, David Card, John DiNardo, Harrison Hong, Lawrence Katz, Morris Kleiner, Robert Mofﬁtt, Jesse
Rothstein, Eric Verhoogen, Hans-Joachim Voth, Wei Xiong, and numerous seminar participants for helpful suggestions. Diane
Alexander, Eric Auerbach, Emily Buchsbaum, Mariana Carrera, Elizabeth Debraggio, Briallen Hopper, Pauline Leung, Sanny Liao,
Stephen Nei, Xiaotong Niu, Zhuan Pei, Andrew Shelton, and Fanyin Zheng provided outstanding research assistance. We gratefully
acknowledge research support from the Center for Economic Policy Studies at Princeton University.
“[L]aymen and economists alike tend, in my view, to exaggerate greatly the extent to which
labor unions affect the structure and level of wage rates.” – Milton Friedman, 19501
“Everyone ‘knows’ that unions raise wages. The questions are how much, under what conditions,
and with what effects on the overall performance of the economy.” – Richard Freeman
and James Medoff, 19842
I Introduction
Over the last several decades in the U.S., there have been important shifts in union membership rates, the
composition of unions, and the frequency and success of organizing drives. In the U.S. the union membership
rate fell from 27 to 13 percent between 1970 and 2000, compared to a decline from 38 to 27 percent in EU
countries during this period (Visser, 2006). This trend in the U.S. masks the even steeper decline in the
private sector from about 25 percent to 9 percent from the early 1970s to 2000 (Farber, 2005).3 Coincident
with this development was a decline in new union orgainizing activity: in 1966 more than 200,000 private
sector workers gained union representation status – achieved through the U.S. system of union recognition
through workplace representation elections – compared to approximately 80,000 in 2006.4
A key to assessing the distributional and productivity implications of these shifts is measuring the extent
to which unionization impacts ﬁrms’ proﬁtability. There is little doubt that employers generally do oppose
unions. An example receiving recent national attention is Wal-Mart’s effort to resist unionization – from
its strategic location of stores in areas less favorable to unions to its hard-line stance against organization
(Basker, 2007). According to a handbook the retailer distributed to its managers, “Staying union free is a
full-time commitment...The commitment to stay union free must exist at all levels of management – from
the Chairperson of the “Board” down to the front-line manager....”5
And the fact that in the U.S. new unionization typically occurs discretely at an employer at a particular
point in time allows one to ﬁnd isolated cases that at ﬁrst blush seem to conﬁrm the fears of employers like
Wal-Mart. For example, in a March 1999 National Labor Relations Board (NLRB) representation election,
1See Friedman (1950).
2See Freeman and Medoff (1984).
3By 2009, the majority of unionized workers in the U.S. were employed in the public sector (U.S. Census Bureau, 2010).
4Based on a tabulation of NLRB election data. This decline has occurred despite a recent increase in the union win rate which
has been trending upward since 1980, reaching 72 percent in 2009 from a low of 42 percent in 1982.
5Quoted in Featherstone (2004).
1
workers at National Linen Service (NLS) Corp., a large linen supplier, voted by an over 2 to 1 margin to
organize as a local chapter of the Union of Needletrades, Industrial, and Textile Employees (UNITE). The
stock market response appeared to punish NLS in a severe, though perhaps not swift, fashion. Figure I
shows the cumulative return of NLS’ stock for the two years prior to and following the election, as well as
the cumulative return of a broad market index over the same period. Before the election, the returns for NLS
and the market tracked each other quite closely. But immediately following the election, NLS began to lag.
By March 2001, the price of NLS shares had fallen by about 15 percent, while the broad market index had
increased by about 25 percent since the election.
But how general is this phenomenon? Is NLS the exception or the rule? Despite an enormous literature
documenting numerous aspects of unions and their role in the labor market, the magnitude of an “average”
effect of unions on ﬁrm performance throughout the economy remains somewhat unclear.
Empirically, there are at least three reasons why measuring these effects is quite challenging. First, largescale
establishment or ﬁrm-level micro-data containing the relevant information on the extent of unionization
are not readily available. Second, even when such data are available, omitted variables and the endogeneity of
unionization at the ﬁrm-level makes it difﬁcult to separate causal effects from other unobserved confounding
factors.6 Third, it is difﬁcult to ﬁnd data that can also be plausibly representative of the population of
unionized companies in the United States.7
Furthermore, from a theoretical standpoint, it is not obvious to what degree unions should affect ﬁrms.
One view, articulated by Friedman (1950), is that workers would reject substantially above-market wages,
knowing full well that such wages could adversely affect job security. Unions, after taking these considerations
into account, would tend to moderate wage demands.8 Moreover, ﬁrms may respond to a unionization
threat by conceding higher wages and better working conditions. Accounting for these forces suggests a
reduction in the gap in compensation and working conditions between union and non-union workforces, at
least in situations where there is a threat of unionization. The possibility that unions may temper their de-
6Hirsch (2007), in a recent study reviewing evidence from ﬁrm- or establishment-level data, suggests drawing inferences from
the existing research with caution, emphasizing omitted variables and the potential endogeneity of union status. Examples of studies
implicitly relying on the assumption that union status is an exogenous variable include the in-depth analyses of Clark (1984), Hirsch
(1991a), and Hirsch (1991b).
7The limited generalizability of many of the studies is a another limitation that Hirsch (2007) emphasizes. For example, the
cement industry is examined in Clark (1980a) and Clark (1980b), hospitals and nursing homes in Allen (1986a), the construction
industry in Allen (1986b), and sawmills in Mitchell and Stone (1992).
8It is this line of reasoning that led to Friedman’s view that the impact on wages was exaggerated (Friedman, 1950). Alternatively,
even if unions raise wages, ﬁrms could respond by skill-upgrading their workforce. To the extent this is possible, negative market
value effects could be moderated. The issue of skill upgrading is discussed in Wessels (1994) and Hirsch (2004).
2
mands because of electoral pressure may help explain the results of DiNardo and Lee (2004), who found
generally small differences in wages, employment, and output between unionized and otherwise comparable
non-unionized workplaces in close representation elections.
In this paper, we ﬁrst assess the extent to which the pattern in Figure I is a generalizable phenomenon,
measuring an average overall effect of private sector unionization among publicly-traded ﬁrms in the U.S. To
do so, we begin with a sample frame that is the universe of all ﬁrms with NLRB union representation elections
between 1961-1999. Since a large number of unionized workplaces in the U.S. come into existence via a
secret-ballot election on the question of representation, this population provides a reasonable representation
of newly unionized workplaces and, to the extent they survive, the future stock of unions in the United States.
We begin analyzing the stock market reaction to union victories using event-study methodologies. The
most distinctive feature of our data – crucial for our research design – is the long panel (up to 48 months
before and after the election) of high-frequency data on stock market returns for each ﬁrm. This feature
allows us to use the pre-event data to test the adequacy of the benchmarks used to predict the counterfactual
returns in the post-event period. The long panel also allows us to examine returns several months beyond
the event, so as to capture the long-run expected effects of new unions, without having to rely heavily on the
assumption that the stock price immediately and instantaneously adjusts to capture the expected presence of
the unions.9
Our event-study analysis reveals substantial losses in market value following a union election victory –
about a 10 percent decline in market value, equivalent to about $40,500 per unionized worker. According
our calculations, if unionization represented a one-to-one transfer from investors to workers through higher
wages, this magnitude would be in line with a union wage premium of 10 percent. Since the total loss of market
value represents the sum of transfers to workers and any other productivity impacts of unionization this
implies, for example, that if the true union compensation premium were greater than 10 percent, there would
be positive productivity effects of unions. The evidence supporting our event-study estimates is compelling:
we ﬁnd that these ﬁrms’ average returns are quite close to the benchmark returns every month leading up to
the election, but precisely at the time of the election, the actual and benchmark returns diverge. The results
for these ﬁrms are robust to a number of different speciﬁcations. In the sample of ﬁrms where we know
that the union is a small fraction of the workforce, we do not ﬁnd a similar divergence of returns from the
9In an earlier version of the paper, we also provided some suggestive evidence on the long-run effects of union victories on
accounting variables found in Compustat data. These results are presented in the Online Appendix.
3
benchmark.
Importantly, we ﬁnd that the effect takes 15 to 18 months to fully materialize, a somewhat slow market
reaction. As we discuss below, this short-run mis-pricing can persist if exploiting the slow reaction is not
sufﬁciently proﬁtable to arbitrageurs. Indeed our own analysis shows that strategies designed to exploit the
mis-pricing entail a signiﬁcant degree of fundamental risk. The fact that union victories are sufﬁciently rare
and spread throughout time prevents the necessary diversiﬁcation that could generate an attractive arbitrage
opportunity. For example, our analysis suggests that attempts to exploit the short-lived mis-pricing would
lead to a portfolio that would be dominated by simple buy-and-hold strategies.
The event-study estimate appears to average a great deal of heterogeneity in the effects. We additionally
employ a Regression Discontinuity (RD) design, implicitly comparing close union victories to close union
losses, and consistent with DiNardo and Lee (2004), we ﬁnd little evidence of a signiﬁcant discontinuous
relationship between the vote share and market returns. If anything, the RD point estimates show a 4 percent
positive (though statistically insigniﬁcant) effect of union certiﬁcation (vis-a-vis union defeat). The eventstudy
estimates vary systematically by the observed vote share, with the largest negative abnormal returns
for cases where the union won the election by a large margin.
We use our estimates to make predictions for the effects of policies that lower the threshold for new
unionization. To do so, while also incorporating unions’ and ﬁrms’ responses to the new policy, requires
modeling their behavior and interactions. We choose as our framework a two-party model of electoral
competition, where the ﬁrm and the union are each seeking to win the sympathies of the “median” voter in
an NLRB election. As is standard in this class of models, despite having opposing interests, the two parties
may be forced to propose a level of compensation (accompanied by a risk of job loss) that is closer to the
preferences of the median voter.
Within this framework, which is reminiscent of Friedman’s view, the RD design estimate of the unionization
effect identiﬁes the gap between the union’s and ﬁrm’s proposals for workplaces where the median
voter has moderate demands. Depending on how aggressively ﬁrms and unions court voters, this gap could
be close to zero, even if on average – including both small and large electoral victories – unions signiﬁcantly
affect the proﬁtability of ﬁrms. Viewed through the lens of this model, the pattern of results imply that
for most union recognitions, the workers – who consider the possible adverse employment consequences to
higher wages – are not particularly demanding. In a smaller share of elections where the effects of a union
win are large, workers have more extreme demands, which are moderated by unions, who place weight on
4
winning elections. Overall, our policy simulation exercise suggests that a policy-induced increase in the win
rate from 33 to 70 percent would lead to a 4.3 percent decline in market value, averaged across all ﬁrms targeted
by unions (including ﬁrms that unionize under the new policy, as well as those that remain nonunion).
For a more dramatic policy that increases the win rate from 33 to nearly 99 percent, the estimate is a decline
of about 11 percent averaged across all ﬁrms targeted by unions for organization.
The remainder of the paper is organized as follows. We provide some institutional details in Section II
that are relevant to our research design, which we describe along with our data. We present and discuss the
empirical results in Section III. In Section IV we present a structural model, which we then use to conduct
counterfactual policy simulations. Section V concludes.
II Institutional Background, Data, and Research Design
The National Labor Relations Act provides the legal framework by which most workers in the United States
become unionized.10 Workers who organize into unions through the procedures speciﬁed by the NLRA
are guaranteed the right to bargain collectively. There are several ways a group of workers may become
unionized under the auspices of the NLRA, though it is believed that most new unionization occurs through
representation elections (Farber and Western, 2001). There are several steps involved in this process, which
are described in detail in DiNardo and Lee (2004). Brieﬂy, when a group of workers decides to organize, they
ﬁrst petition the NLRB to hold a representation election. To be legally granted an election, the petition must
be signed by at least 30 percent of the workforce, typically over no longer than a six month period. Once the
NLRB determines the appropriate bargaining unit, it holds an election at the work site. The union wins the
election with a simple majority of support amongst the workers. Barring objections by the employer, a win
means the union is certiﬁed as the exclusive bargaining agent for the unit and that the employer is legally
required to bargain with the union in good faith.
Our research design and subsequent data collection were motivated by our desire to estimate the average
effect of union victories and losses in representation elections on ﬁrm market value, and to attempt to address
some of the aforementioned puzzles and challenges in the literature. In collecting the data our goal was to
obtain information on the proﬁtability of ﬁrms over a long time span, with a panel structure allowing for an
event-study design with a long event window. Our sample size needed to be large enough so we could also
10Exempt from the NLRA are state and local workers, who are covered by state collective bargaining laws, and railway and airline
workers, who are covered separately by the Railway Labor Act (RLA).
5
estimate the cross-sectional relationship between post-event abnormal returns and the union vote share. For
these reasons, and because we were also interested in how the union effect evolved over time, we sought to
collect information on elections over as many years as possible. Since data on the proﬁts of privately held
ﬁrms are difﬁcult to come by, we focused on publicly traded ﬁrms for which stock market information and
other performance measures are available through mandatory disclosure.
II.A Data Set Assembly
This study primarily uses three sources of data: election results from the NLRB, data from the Center for
Research on Security Prices (CRSP), and the CRSP/Compustat Industrial Quarterly Merged Database.
The NLRB began publicly reporting representation election vote tallies in 1961. However, previous
studies using NLRB election data typically used records that were already in electronic form (e.g. Farber
and Western, 2001; DiNardo and Lee, 2004; and Holmes, 2006). We use those data for the 1977-1999 period,
but augment those with data from 1961-1976 that we digitized for this study.11 Data for the 1961-1976 period
were hand-entered from hard copies of NLRB monthly election reports. Among other things, the NLRB
data set contains the number of voters who voted in favor of the union, the number of voters voting against
the union, the number of eligible voters, the name of the company, a two digit industry code, the city and
state of the election, and the month that the NLRB closed the election.12 The CRSP and Compustat data
were obtained from Wharton Research Data Services.
The primary objective of the data assembly process was to match companies in the NLRB election ﬁles
to companies in the CRSP data ﬁle. The procedure for matching establishments in the NLRB dataset to ﬁrms
in the CRSP dataset is detailed in the Data Appendix. This matching process is complex because while the
NLRB ﬁle provides the company name where the election took place, most other identifying information is
unknown.13 However, as explained in the Appendix, we are conﬁdent that the match is high quality.
Previous event studies of representation elections use samples of elections with a very large number of
eligible voters. Ruback and Zimmerman (1984) and Bronars and Deere (1990) limit their sample to elections
11The 1977-1999 period data were obtained from Thomas Holmes’ website (http://www.econ.umn.edu/~holmes/data/geo_spill/)
and are used in Holmes (2006).
12For a limited number of years the NLRB data has information on the calendar date of the election and the calendar date the
NLRB closed the case.
13The location of the election is not very useful for matching because the CRSP ﬁle only contains the location of company
headquarters, which may differ from the location of any establishment undergoing a recognition election. The only additional
information that could help us identify a match is the two digit SIC industry code of the establishment. However, the industry of
an establishment may differ from the primary industry of the ﬁrm. This variable is more useful as a check for the validity of the
matches.
6
with at least 750 eligible voters. Elections of this size are quite rare, thereby resulting in small sample sizes
(54 union victories in the main sample of Ruback and Zimmerman, 1984). We believe that the effects of
these elections are easier to detect if the number of eligible voters is large relative to the size of the ﬁrm.
However, limiting the sample to large elections is neither necessary nor sufﬁcient to achieve this objective.
Because many of these elections take place in very large ﬁrms, the ratio of voters to total ﬁrm employment is
no larger here than for moderately sized elections. While we do not have the exact sample used by Ruback
and Zimmerman (1984), we can attempt to replicate it based on their description of the sample selection
scheme.14 Using their sample selection scheme we ﬁnd that in more than 10 percent of the elections, less
than 1 percent of the ﬁrm’s workforce voted. In our reproduction of their sample, the median percentage
of the workforce voting in an election is 5 percent.15 By contrast, our main analysis limits the sample to
elections where at least 5 percent of the total workforce voted.16 The median election in our sample consists
of 13 percent of the company’s workforce voting (mean = 22 percent).17 Therefore, our sample selection
scheme not only provides us with elections that are relatively salient for a given ﬁrm (or, at a minimum,
excludes those elections which are clearly not salient), but also yields a substantially larger sample size
compared to what we would have obtained using the Ruback and Zimmerman (1984) criterion. Our baseline
sample is almost eight times larger than the Ruback and Zimmerman (1984) sample.
We present summary statistics of ﬁrm characteristics in Table I. Columns (1) and (2) correspond to
elections where at least 5 percent of the workforce voted (hereafter the “≥ 5% sample”) for UV (“Union
Victory”) and UL (“Union Loss”) ﬁrms respectively. Columns (3) and (4) correspond to elections where
less than 5 percent of the workforce voted (hereafter the “< 5% sample”) for UV and UL ﬁrms respectively.
Looking at the ﬁrst row of Table I, there are about twice as many elections in the < 5% sample than in
the ≥ 5% sample, and in both samples there are about twice as many ﬁrms where the union lost than where
the union won. Not surprisingly, ﬁrms in the ≥ 5% sample tend to be substantially smaller than ﬁrms in
the < 5% sample. This inference can be made by comparing a variety of measures, including employment
14Using the Ruback and Zimmerman procedure we ended up with almost twice as many elections as they had considered over
the same time period. The only information that Ruback and Zimmerman had that we do not is the petition date. They excluded
elections where the petition date was unavailable. We therefore infer that this exclusion restriction would have resulted in us dropping
50 percent of the elections in the sample.
15Huth and MacDonald (1990) conduct an event-study of decertiﬁcation elections. Their sample selection scheme involves all
decertiﬁcation elections involving at least 250 workers between June 1977 and May 1987. They also do not condition on there being
a sufﬁciently high fraction of a ﬁrm’s workers involved in the election. Our (inexact) reproduction of their sample has a median
fraction of the workplace voting of 2 percent, with approximately 30 percent of elections in the sample involving less than 1 percent
of the company’s workforce.
16Total employment in the year of the election is from the Compustat annual ﬁles.
17We do not use elections where employment information is missing.
7
(3,541 vs.73,223 employees) and market value ($338 million vs. $5.9 billion in 1998 dollars, using the
more broadly available CRSP measure). However, the ≥ 5% sample corresponds to bigger elections, with
an average of 453 workers voting as compared to an average of 291 in the < 5% sample. Table I also shows
the delisting rate for companies. We report the fraction of companies delisted in the two years before or
after the election. UV ﬁrms are slightly more likely to delist than UL ﬁrms (10 versus 8 percent delisting
rates respectively).18 While this difference is not large, we will consider several approaches to address this
issue, as well as the presence of missing returns more generally. These approaches involve imputing missing
returns, estimating all models excluding periods with missing returns, or limiting the sample to ﬁrms that
have no missing returns in the event window. Simply excluding missing values has the disadvantage that
some of the changes in cumulative returns over time may reﬂect ﬁrms that are entering or dropping out of the
sample. Using a balanced panel has the advantage that we can be sure that any differences over time are not
caused by compositional differences. However, a balanced panel does involve discarding a large number of
elections and implies that inclusion into the sample may depend on the realization of the dependent variable.
We will demonstrate that the results are not sensitive to the approach employed.
II.B The Event-Study Method
Our objective is to assess the impact of union elections on the stock market value of ﬁrms. Ideally, we would
like to compare the ﬁrm’s stock returns to the returns the ﬁrm would have experienced in the absence of a
union organizing event. The event-study method provides a framework for estimating this counterfactual
return.
As is standard in the ﬁnancial economics literature, we deﬁne the abnormal return as the difference
between a stock’s actual return and the expected return given market conditions. For the company corresponding
to union representation election i, in month t, the abnormal return is:
ARit ≡ rit −E[rit|Xt]
where rit is the actual return and E[rit|Xt] is the predicted return. For this study, rit is the CRSP monthly
holding-period return including distributions, which is constructed using prices that are adjusted for splits
and distributions.19
18We deﬁne delisting as any company with a non-missing delisting return in the CRSP dataset.
19When stocks are delisted we use CRSP delisting returns. We replace missing returns with the predicted return (E[rit|Xt]) to
8
For convenience, we express time in terms of months relative to the event:
ARiτ ≡ riτ −E[riτ|Xτ]
where ARiτ is the abnormal return of the security corresponding to election i in the τ’th month relative to
the event.
Because returns of companies with unionization events may vary systematically before the elections, perhaps
due to anticipation of the event, and because the market may not react instantaneously, we are interested
in the cumulative abnormal return (CAR) in a window surrounding the election. The CAR corresponding to
event i between months T1 and T2 relative to the event is:
CAR(T1,T2)i ≡
T2
∑
τ=T1
ARiτ
The statistic of interest is the average (across N ﬁrms in the sample) cumulative abnormal return:
ACAR(T1,T2) ≡
1
N
N
∑
i=1
CAR(T1,T2)i
We will present the average cumulative abnormal return for the set of union victory (UV) and union loss
(UL) ﬁrms beginning two years prior to the election. Our decision to use such a long event window is in
part the consequence of having information on the month that the NLRB closed the case, rather than the
exact calendar date. By considering a very long pre-event window we can verify that any difference in the
cumulative return of the UL and UV ﬁrms and any counterfactual (or “benchmark”) portfolio is not simply a
continuation of differential pre-event trends. If there are signiﬁcant departures between our predicted returns
and the observed returns over the two year period before the event, we consider any estimates obtained from
the post-event data to be invalid.20 This approach is a direct application of conventional testing of overidentifying
restrictions for “difference-in-difference” modeling in labor economics program evaluation.21
The long panel also allows us to examine returns in the months beyond the event, so as to capture the
mitigate survivorship bias, though the results are not sensitive to how missing values are treated. Speciﬁcally, the results are not
sensitive to simply ignoring missing values, nor to only selecting companies with no missing returns in the entire event-period.
20An alternative interpretation of pre-election divergence in the predicted and actual returns is the diffusion of anticipatory information
regarding the election outcome. Recognizing this alternative, we allow for non-zero excess returns in a short window prior
to the event, but conclude that any signiﬁcant divergence over a long-period of time prior to the event is evidence of a mis-speciﬁed
model.
21For example, see Ashenfelter and Card (1982) and Heckman and Hotz (1989).
9
long-run expected costs to the ﬁrm without having to rely on the assumption that the stock price immediately
and instantaneously adjusts to the presence of the union. Note that in typical event studies, T1 and T2 usually
indicate days relative to the event, but since in our study we are looking at long-run trends, T1 and T2 denote
months relative to the unionization event.
A critical decision in event-studies is how to model E[rit|Xt]. A common approach for computing abnormal
returns in long-run event-studies involves the use of reference or “benchmark” portfolios matched on
a ﬁrm’s characteristics (see Barber and John D. Lyons, 1997; Lyon et al., 1999; and Brav, 2000). The advantages
of this approach are that the benchmark can be constructed in-sample and that it allows for shocks
occurring by chance that affect ﬁrms with similar characteristics. We employ this approach, matching every
ﬁrm in our sample to a portfolio of ﬁrms in the same size-decile.22 As a probe for robustness we have also
used the CRSP equally-weighted NYSE/AMEX/NASDAQ index as a benchmark, comparing ﬁrms both in
the same size decile and in the same one-digit SIC industry.23
A complication arises when trying to deﬁne the “event.” The appropriate event is the date on which
most of the information on the probability of future unionization is incorporated. For much of the sample
(1961-1976) we only observe the month that the NLRB closed the case. While we have a well-deﬁned event,
it is not the only relevant event and it may not be the most important one. Alternatively, potentially important
events are the petition and election dates. Using post-1977 data, where both the election and case closure
calendar dates are available, we ﬁnd that the median time between the election and NLRB case closure is ten
days. In some cases, typically when one of the parties issues a challenge, this gap can be considerably longer.
In 5 percent of the elections it took at least six months for the NLRB to close the case. While we do not have
data on when the petition was submitted to the employer, it is known from Roomkin and Block (1981) that
elections usually occur very soon after the petition. In their sample, 42 percent of elections occurred within
one month of petition and 83 percent within two months. Therefore, we do not believe that using the month
the NLRB closed the election presents serious problems for estimation if most of the new information is
revealed at or after the petition date. To assess whether gradual diffusion of news led to abnormal returns
prior to the closing date it is useful to examine a long pre-event window. We believe, however, that it will be
22CRSP produces indices for such purposes. Speciﬁcally, every year CRSP allocates companies into one of ten size deciles, based
on market-value. The value-weighted average return of securities in these deciles are then calculated on a monthly basis. CRSP also
produces a cross-walk that allows one to link each security to the appropriate size decile.
23We cannot match on the book-to-market equity ratio, as many studies do, because this variable is unavailable for a large number
of companies in our sample, especially in the earlier periods. We also used the calendar time portfolio approach developed by Jaffe
(1974) and Mandelker (1974) and advocated by Fama (1998). We ﬁnd qualitatively similar results from this analysis, as shown in
the Online Appendix.
10
difﬁcult to empirically distinguish the market’s anticipation of unionization from an inadequate comparison
portfolio.
The event-study method can inform us on how the equity value of ﬁrms responds to certiﬁcation elections.
We can also estimate event-study models for elections with varying degrees of union support to explore
heterogeneity in the effect size. A more complete investigation of heterogeneity in the impact of certiﬁcation
elections on stock market performance involves estimating the post-event cumulative abnormal return
for every election and relating these to the vote share in a ﬂexible way. We conduct this analysis to examine
the heterogeneity in the stock market reaction to election outcomes and to determine whether there
is a discontinuous relationship between cumulative abnormal returns and the vote share at the 50 percent
threshold.
III Empirical Results
III.A Event-Study Estimates
In Figure II we plot the average cumulative return of union victory ﬁrms against the average cumulative
return of the size-matched reference portfolios over the same time period.24 The ﬁgure reveals that both UV
ﬁrms and the corresponding reference portfolios have almost identical trends in returns prior to the union
victory. However, near the time of the election there is a pronounced downward break in the returns of UV
ﬁrms relative to the benchmark, persisting for approximately a year and a half. The average cumulative
abnormal return implied by this divergence is approximately -10 percent.
The pattern we ﬁnd contrasts with that reported in the well-known study of Ruback and Zimmerman
(1984), which also examines the stock market reaction to NLRB union certiﬁcation events.25 Speciﬁcally,
given their sample selection scheme (as described above), their data show substantial negative abnormal
returns that emerge well before the unionization event: speciﬁcally, a decline in market value of about 7
percent between the 12th and 7th months preceding unionization. This pattern raises the question of whether
the post-election decline in the stock market valuation that they ﬁnd – a 3.8 percent drop within a few months
24For convenience, we will often refer to the event month as the “election month," though it should be understood that we actually
only know when the NLRB closed the case.
25There are a number of other studies that examine various aspects of unions through stock market reactions. They typically
do not aim to generate effects of unionization (versus the absence of unions), as they use samples of already unionized ﬁrms or
industries. See Abowd (1989), Becker and Olson (1986), Neumann (1990), DiNardo and Hallock (2002), and Becker (1987). Olson
and Becker (1990) is an exception in this regard, as it examines the impact of the passage of the National Labor Relations Act on 75
ﬁrms that were at risk of being unionized in the 1930s.
11
surrounding the unionization event – reﬂects unionization or the factors which led to the pre-election trend
in the ﬁrst place.26 While Ruback and Zimmerman (1984) have no explanation for this signiﬁcant decline,
they argue that it is unlikely to indicate anticipation of the outcome of the election due to its timing.27 The
issue of an absence of solid evidence of comparable trends prior to the event has arisen in other “differencein-difference”
analyses using establishment-level plant data, such as in Lalonde et al. (1996) and Freeman
and Kleiner (1990b).
To assess the magnitudes and statistical signiﬁcance of the effect implied by Figure II, in Figure III we
plot ACAR(−24,τ), for τ = −24 through τ = 24, with 95 percent point-wise conﬁdence intervals. Note
again that τ denotes the number of months relative to the election event. The ﬁgure shows that the downward
shift in abnormal returns emerging soon after NLRB case closure is statistically signiﬁcant. Accumulating
the effects starting at time zero, we can reject the null hypothesis that the average abnormal returns are equal
to zero ﬁve months after the event at a 5 percent level of signiﬁcance.28 We interpret Figures II and III as
providing evidence that union election wins correspond to large negative abnormal returns.
Figure IV contains the plot of the average cumulative return for union loss ﬁrms against the average
cumulative return of the size-matched reference portfolios. As with the UV ﬁrms, the reference portfolios
closely track the progression of UL ﬁrms prior to the election, but unlike UV ﬁrms, the returns of UL ﬁrms
do not diverge from the benchmark after NLRB case closure. If anything, there is a moderate increase in the
cumulative return of UL ﬁrms relative to the benchmark, though in Figure V, which presents the difference in
these series with conﬁdence bands, we see this increase is not statistically signiﬁcant at conventional levels.29
We have conducted a variety of analyses to determine whether the patterns seen in Figure II and Figure
IV are robust. These analyses include: not imputing missing returns (Online Appendix Figures VI and
XII); using a balanced panel (Online Appendix Figures VII and XIII); excluding elections where cumulative
abnormal returns following case closure are less than or equal to the 5th percentile or greater than or equal
to the 95th percentile of all post-event cumulative returns (Online Appendix Figures VIII and XIV); using
26Speciﬁcally, the main estimate of -3.84 percent is computed by taking the one-month change associated with the petition date
and adding it to the one-month change associated with the date of the actual certiﬁcation. This can be seen as the summation of the
third and ﬁfth rows, which equals the ﬁrst row of the third column in their Table II. Their main estimate can also be seen in their
Figure I(c) as the summation of the two downward notches around the petition and certiﬁcation dates.
27Speciﬁcally, on p.1145, they note that “[t]he abnormal return for these ﬁrms in the 6 months immediately preceding the petition
is 0.16 percent. This timing suggests that the pre-petition abnormal returns are not due to unionization. Instead, the results suggest
that ﬁrms in which unions are successful experienced declines in value prior to the union activity.”
28We also compute ACAR(0,τ) from τ = 0 to τ = 24, and obtain a point estimate (and standard error) of −0.092 (0.033) for
ACAR(0,24).
29Our more precise estimate from computing ACAR(0,24) yields a point estimate (standard error) of 0.029 (0.028).
12
a four year pre-event window (Online Appendix Figures IX and XV); using an industry×size matchedreference
portfolio (Online Appendix Figures X and XVI); using the CRSP equally-weighted market index
as the reference portfolio (Online Appendix Figures XI and XVII), and taking into account the fact that
multiple election events may have occurred in the same ﬁrm (Online Appendix Figure XVIII)30. In all cases
the overall pattern of cumulative returns look very similar to those seen in Figures II and IV.31
Table II, Panel A presents average cumulative abnormal returns following union victories. The ﬁrst column
corresponds to the use of the size-matched benchmark. Column (2) corresponds to the industry ×sizematched
benchmark. Column (3) corresponds to the CRSP equally-weighted NYSE/AMEX/NASDAQ index
benchmark. In the ﬁrst row of Panel A we report ACAR(0,24) for each of the three benchmarks. The estimated
post-election average cumulative abnormal returns range from -9 to -10 percent and are signiﬁcant
at the 1 percent level. Standard errors are calculated using a cluster-robust variance estimator proposed by
Cameron et al. (2006). To gauge magnitudes, we calculate that a 10 percent negative return corresponds
to approximately $20 million in lost market value (in 1998 dollars). We then divide this ﬁgure by the total
number of workers who were eligible to vote in these ﬁrms, which yields a ﬁgure of $40,522 per newly
unionized worker.32 Suppose we take annual income of workers prior to unionization to be $25,000 (in 1998
dollars).33 Assuming that future earnings for the ﬁrm fall dollar for dollar with increases in wages – and
suppose the union wage premium was 10 percent – then 10 percent of $25,000 in perpetuity, at a 6 percent
discount rate, yields $41,667 in discounted value, which is roughly equivalent to our estimate of $40,522.34
30In our main analysis, we abstract from the occurrence of multiple elections within a 4-year interval at the same ﬁrm by assuming
each election and the associated ﬁrm market values are separate. We simply regress the monthly abnormal returns on a set of 49
“event-time” dummy variables with the speciﬁcation E [ARit] = ∑24
τ=−24 Dτ
itγτ, where Dτ
it is, for ﬁrm i in month t, equal to 1 if the
election event occurred at time t −τ, and 0 otherwise. In this basic speciﬁcation, for any month-ﬁrm observation, only one of the
event-time dummy variables is equal to 1. In fact, out of the 414 union victory elections examined, 126 elections occur within
4-years of another election at the same ﬁrm. To gauge the importance of this, we use the same sample and regression, but allow
more than one of the event-time dummy variables to equal 1 for any month-ﬁrm observation for which more than one election
occurred within a 4-year period. This implies an additive effect of multiple elections: if an election occurred at time t′ and t′′, then
E [ARit] is equal to γt−t′ + γt−t′′ . Online Appendix Figure XVIII shows the results from this speciﬁcation, accumulating the γs to
form estimates of ACAR(−24,τ) from τ = −24 to 24. The effects are slightly larger in magnitude, but overall the pattern is very
similar to that of Figure III. We also examined a sub-sample of 347 “ﬁrst elections”, and ﬁnd a similar pattern.
31A possible exception is Online Appendix Figure XV, which shows that UL ﬁrms experienced a period of positive abnormal
returns three years before the election. Since much of the prior literature focuses on manufacturing ﬁrms, we conducted the analysis
separately for manufacturing ﬁrms and non-manufacturing ﬁrms. 323 of the 414 elections are at manufacturing ﬁrms; the pattern
and magnitude of effects for this subset mirror that of Figure II. The pattern is less clear for the remaining 91 non-manufacturing
ﬁrms, due to increased sampling variability.
32Here we are taking the number of eligible voters as an approximation to the size of the collective bargaining unit.
33In 1980 (the mid-point of our sample frame) the average non union wage was $12.43 in 1998 dollars (Hirsch and Macpherson,
2008), translating to approximately $25,000 in annual income.
34Our 10 percent wage premium is on the low side of union/non-union differentials from conventional cross-sectional wage
regressions using household survey micro-data. Blanchﬂower and Bryson (2007) report adjusted union wage gap estimates for the
private sector that range from 12.7 to 22.4 percent in the period between 1973 and 2002.
13
This appears to be a plausible value. It is important to note that this ﬁgure – which is based on the impact
of union recognition – averages the effects for when the union secures a ﬁrst contract and when they do
not. If one assumes that the effect is smaller for ﬁrms where the union does not secure a contract, then our
estimates are a lower bound for the magnitude of the effect of a union victory and a contract.35 Of course,
we are unable to say whether the loss in equity value reﬂects increases in wages, beneﬁts, or inefﬁciencies.
Additionally, if unionization leads to an increase in productivity, then 10% may be an underestimate of the
actual compensation premium.
In the second row of Table II we report ACAR(-24,-4), the average cumulative abnormal return prior
to case closure, excluding the three months immediately preceding the event. ACAR(-24,-4) is statistically
indistinguishable from zero in all three speciﬁcations. The lack of signiﬁcant abnormal returns prior to the
election indicates that the market did not anticipate these events, on average, and also suggests that all three
benchmarks do a reasonable job of predicting average returns of the portfolio of UV ﬁrms. Table II, Panel
B reports the same set of estimates for union loss ﬁrms. Consistent with what we observe in Figure V, the
cumulative abnormal returns are close to zero and statistically insigniﬁcant.
One possible concern is that elections are endogenous to the performance of ﬁrms. However, we ﬁnd
little evidence that this is the case. The ﬁrms in our sample track their benchmarks quite closely prior to the
election, so it does not appear to be the case that the election is a result of the ﬁrms under- or over-performing
the benchmark. There is also no indication that the ﬁrm’s performance in the two years prior to the election
is systematically related to how the union fares in the election. This can be seen in a number of ways. For
example, looking at Figure II, winners and the benchmark portfolio are not trending differentially prior to the
election. To test this hypothesis more directly we have regressed the union vote share in the election on the
cumulative abnormal return from -24 to -4 and found no signiﬁcant relationship between the two variables.36
If workers are deciding on the performance of the ﬁrm, they are basing their decision on forecasts of future
performance rather than past performance. While we cannot rule out this possibility, it is not obvious how
workers could forecast future share prices of the ﬁrm, and why it would be optimal for them to ignore past
performance. Moreover, it is not clear why it would be optimal to unionize when the ﬁrm is projected to
perform poorly.
Our sample selection scheme was partly predicated on choosing elections where a sizable fraction of the
35Cooke (1985) estimates that 25% of union election victories do not lead to contracts.
36Speciﬁcally, we estimate a coefﬁcient of -0.006 with a standard error of 0.09. This estimate is not sensitive to the pre-event
window over which the CAR is calculated.
14
ﬁrm’s workforce was voting: in practice we used a 5 percent cutoff. As a falsiﬁcation exercise we examine
elections where a small fraction of the ﬁrm’s total workforce voted. The idea is that we should not see effects
in ﬁrms where only a very small share of the employees voted. In Table III we examine whether cumulative
abnormal returns following an election become more pronounced when a larger share of the ﬁrm’s workforce
is participating. Speciﬁcally, using the full sample of elections we relate ACAR(0,24)i, where i denotes an
election, to the share of the ﬁrm’s total workforce in the bargaining unit. As seen in Column (1), when the
union wins the election and the fraction of the ﬁrm’s workforce in the bargaining unit is essentially zero,
the ﬁrm experiences a small and positive abnormal return. As we would expect, as the share of the ﬁrm
involved in the election increases, the resulting effect on the abnormal return becomes more pronounced.
Each percentage point increase in the share of the ﬁrm’s employees voting in the election is associated with
a third of a percentage point decline in the post-event cumulative abnormal return, a relationship which is
statistically signiﬁcant at the 5 percent level. Column (2) presents these estimates for the union loss sample.
The negative relationship in the post-event cumulative abnormal return and the share of the workforce voting
is not present. In fact, there is a positive relationship, which is what we would expect if union losses resulted
in positive abnormal returns.
III.B Discussion of the Results and Additional Analyses
Speed of Adjustment
We now turn to a important feature of Figure III: the relatively slow emergence of the effect, with an
abnormal return beginning around the time of the election and growing for approximately 15 months. The
pattern from our event study reveals that even if investors in each individual UV ﬁrm believe their forecasts
for future earnings are unbiased, immediately following the election, these investors are, as a whole, systematically
underpredicting the eventual value implications of unionization. It is widely understood that there
exist irrational or misinformed investors, whose behavior can potentially push prices away from fundamental
value.37 The real puzzle, on the face of it, is how market forces would allow this implicit and systematic
under-prediction to persist over such a long period. After all, viewing the group of UV ﬁrms as a portfolio,
investors could attempt to take advantage of the forecastable delayed reaction, which would exert downward
37As an example (and one of many different possibilites) of such uninformed behavior, it is easy to imagine that UV ﬁrms are
not being evaluated as a portfolio, but instead individually monitored. Given that immediately upon a union victory, there ensues
a period (of potentially several months) of uncertainty as to the signing and terms of a ﬁrst contract, there is room for investors to
believe that the UV ﬁrm they are holding will perform better than the UV ﬁrms ultimately do on average.
15
pressure on the UV stock prices immediately after the union victory, turning the slow reaction into a quick
one. How can this apparently slow reaction of the stock market persist in the long-run? We consider four
possible explanations.
First, it is possible that stock prices exhibit momentum because information, especially negative information,
diffuses gradually to investors, as suggested by Hong et al. (2000b). We apply the approach used
in Hong et al. (2000b) to our sample, and compare ﬁrms with and without analyst coverage. According to
I/B/E/S International analyst data, only 50 percent of the ﬁrms in our sample had analyst coverage at the time
of the election, meaning that these elections may not have been widely publicized or followed.38 In Figure VI
we compare average cumulative abnormal returns for companies that did and did not have analyst coverage
at the time of the election. Companies with analyst coverage appear to have experienced negative abnormal
returns earlier than those without analyst coverage. But even these experienced a relatively slow-reaction to
the event on average, suggesting that the lack of analyst coverage is not the complete story.39
Second, we consider the possibility that the reaction is actually becoming swifter over time, as the implications
of union victories for market value are becoming more widely known and exploited by investors.
In the Online Appendix, we compare the patterns of average cumulative abnormal return of UV ﬁrms for
elections occurring in the 1961-1983 period to those occurring in the 1984-1999 period. The analysis shows
that the average effect of a union certiﬁcation win on ﬁrm performance exhibits a fairly similar pattern over
the two time frames; the speed of the reaction does not seem to be increasing over time.40
A third possibility is that the pattern in Figure III is reﬂecting a structural change in systematic risk due
to unionization, so that after adjusting for this change, there is a more precipitous post-event decline, or
perhaps even a small or no decline overall. This potential explanation is analogous to the notion of shifting
betas as an explanation for the well-known post-earnings announcement drift (Bernard and Thomas, 1989).
38The 50 percent ﬁgure is derived from I/B/E/S International analyst data for years 1976-1999.
39We are aware that companies not appearing in I/B/E/S may still have analyst coverage. This kind of misclassiﬁcation tends
to reduce the measured difference in excess returns between these two groups of ﬁrms, if in fact there are actual differences. It
is unlikely that this measurement problem will affect the relatively slow speed of adjustment for companies covered by analysts,
as these are presumably measured correctly, meaning that our basic conclusion–that analyst-covered companies exhibit a relatively
slow speed of adjustment–still holds.
40We also note that the magnitude of the effect also does not appear to be declining over time, casting some doubt on the notion
that the small union effects found in the DiNardo and Lee (2004) sample (comprised of only post-1984 elections) are due to unions
having weaker bargaining power in the post-1984 period. We have also compared the effects for states with and without right-towork
laws. Conditional on a union winning its election, the stock-market effects of unionization tend to be more pronounced in states
with right-to-work laws than those without. The result is broadly inconsistent with the notion that right-to-work laws fundamentally
weaken unions because of a potential free-riding problem. Farber (1984) and Moore and Newman (1985) suggest that right-to-work
laws are primarily symbolic, reﬂecting a taste against union representation rather than having any real effect, though Ellwood and
Fine (1987) ﬁnd that these laws do decrease union organizing.
16
Using an approach similar to that employed in Bernard and Thomas (1989), we estimate a standard CAPM
regression with our sample: we regress ﬁrm returns on the market return (both monthly, net of the treasury
rate), dummy variables for event time -24, through +24, and interactions of the market return with dummy
variables for eight 6-month periods within the -24 to 24 interval.41 This speciﬁcation allows for betas that
change at 6-month intervals, and month-speciﬁc “Jensen’s alpha”s, which is meant to reﬂect corresponding
risk-adjusted returns. The eight separate point estimates of beta range from 0.99 (s.e. 0.061) for the month
19 to 24 period, to as high as 1.11 (s.e. 0.045) for the -18 to -13 period, with a standard F-test failing to
reject the equality of the betas (p-value of 0.66).42 Importantly, as shown in Online Appendix Figure XIX,
the evolution of the implied cumulative abnormal returns – the running summation of the “Jensen’s Alpha”s
starting at month -24 – is quite similar to that found in Figure III, with a comparable speed of decline and
overall effect size. It appears that a shift in betas is unlikely to explain the pattern of our results.
Finally, we assess a fundamental premise behind the expectation of a swift market reaction to union victories
– that exploiting the slow reaction would be sufﬁciently proﬁtable to arbitrageurs to lead to a correction
of the short-run mis-pricing. As Barberis and Thaler (2003) point out in their survey of behavioral ﬁnance,
“straightforward-sounding textbook arbitrage” differs from real-world arbitrage, as the latter involves potentially
important risks and costs: once they are acknowledged, then predictable mispricing can persist without
it being an attractive arbitrage opportunity. In our context, the question is to what extent would an arbitrageur
– armed with our empirical evidence – consider it an attractive investment opportunity to take advantage of
the slow market reaction to union victories? If such opportunities – after appropriately considering their
risks – are very attractive, then the gradual emergence of the effect, as in Figure III, would indeed remain a
puzzle.
Our analysis suggests, however, the opposite: taking advantage of the slow market reaction is considerably
risky, and not particularly attractive. We show this in two different ways. First, we consider the
individual performance of 414 separate investors – one for each of the ﬁrms in our main UV sample – who
adopt a zero-investment strategy, taking a short position in the UV ﬁrm, and an equal-value long position in
the corresponding benchmark portfolio, upon the month of the election. Panel A of Table IV shows that at 5
months after the event, this “arbitrage” opportunity achieves positive returns for only 61 percent of these hypothetical
investors. That proportion rises to .63 and .65 for 10 month and 15 month horizons, respectively.
41Event month zero is included in the 1 through 6 interval.
42An alternative speciﬁcation, allowing for only two separate betas (pre-event and post-event) yields point estimates (standard
errors) of beta of 1.07 (0.03) and 1.03 (0.03), respectively.
17
As a comparison, if the 414 investors each followed a simple “buy (CRSP NYSE/AMEX/NASDAQ index)
and hold” strategy during the same time frames, returns would be positive for 66, 74, and 77 percent of them
over the 5-, 10-, and 15-month horizons, respectively. As seen in the ﬁrst and third columns of Panel A,
the ratio of the mean to the standard deviation of returns is substantially lower for the “arbitrage” strategy,
compared to a simple “buy and hold” portfolio. To the individual investor closely following a particular ﬁrm
experiencing a successful organizing drive, the knowledge that on average UV ﬁrms experience a delayed
negative price reaction is not particularly helpful.
Second, we consider a single investor who attempts to take advantage of the pattern in Figure III, while
hoping to diversify the portfolio to minimize the risk discussed above. Here, we imagine this investor pursues
a zero-investment strategy, with a short position in a “UV portfolio” and a long position in a “UV-benchmark
portfolio.” The “UV portfolio” is one that consists of stocks (equally value-weighted) of all of the UV ﬁrms
– at each point in calendar time – that are within a 15-month window subsequent to a union election, and
is continuously rebalanced in this way throughout the sample period. The “UV-benchmark portfolio” is
constructed identically, but instead using the UV ﬁrms’ corresponding benchmark portfolios. Once again,
we see that this “arbitrage” strategy is not particularly attractive. Using all of the possible starting months
within our sample period, the second column of Panel B shows that even at 3-year horizons, only two-thirds
of the time would the strategy lead to positive returns; this is compared to 88 percent for a buy-and-holdthe-market
strategy. At every horizon that we examine, the diversiﬁed “arbitrage” strategy is dominated by
“buy-and-hold” both in terms of the fraction of the time there are positive returns, and in terms of the ratio
of the mean to the standard deviation of returns. As Shleifer and Vishny (1997) argue, returns with this kind
of volatility are likely to be avoided by arbitrageurs.43
Overall, Table IV demonstrates that taking advantage of this short-run mispricing falls well short of
delivering riskless proﬁts. Our context seems to satisfy conditions summarized by Barberis and Thaler (2003)
for there to be limits to arbitrage, and hence for mispricing to persist: 1) it is unlikely that the comparison
portfolio (e.g. size-matched ﬁrms) acts as a perfect substitute to the UV ﬁrm for completely eliminating
fundamental risk, and 2) it is difﬁcult to diversify this risk at any given point in time, as illustrated by the
small improvement from Panel A to Panel B in Table IV, perhaps unsurprising given that there are only
43We can also compute the fraction of the time that this arbitrage strategy would beat the index benchmark. At 15, 36 and 60
month horizons, this diversiﬁed UV strategy would beat the index benchmark 31, 29, and 19 perent of the time, respectively. By
comparison, Shleifer and Vishny (1997) report that the odds of a conventional arbitrage of the widely studied and documented
glamour-value anomaly “outperforming the S&P500 index over one year have been only 60 percent”, while “over 5 years the
superior performance has been much more likely.”
18
414 UV events spread across approximately 40 years.44 As Barberis and Thaler (2003) point out, these
conditions, along with risk averse arbitrageurs, would allow security prices to persistently deviate from
fundamental value.45 Furthermore, this simple story ignores the impact of noise trader risk (De Long et
al., 1990) (investors overly optimistic in UV ﬁrms becoming even more optimistic in the short run), and
implementation costs (constraints on short-selling), both of which would further reduce the attractiveness of
exploiting the mispricing.46
It is important to note that our ﬁnding of a market under-reaction to a seemingly important event, is
not particularly anomalous, viewed in the context of many studies in empirical ﬁnance. Systematic underreactions
have been reported in response to IPOs and SEOs (Loughran and Ritter, 1995), mergers (Asquith,
1983; Mitchell and Stafford, 2000), stock splits (Ikenberry et al., 1996), share repurchases (Mitchell and
Stafford, 2000), exchange listings (Dharan and Ikenberry, 1995), dividend initiations (Michaely et al., 1995),
spin-offs (Cusatis et al., 1993), earnings announcements (Ball and Brown, 1968), and predictable changes in
demographics (Dellavigna and Pollet, 2007).47 Indeed, in further exploring the proﬁtability of momentum
strategies documented in Jegadeesh and Titman (1993), Hong et al. (2000a) show that the cumulative returns
of a portfolio that holds a long position in past “winners” and short position on past “losers” grows gradually
– following a similar time pattern to our Figures – and only ﬂattens out after 10 to 24 months.
Compustat Analysis
The results presented up to this point suggest that union victories are associated with negative abnormal
returns. In the Online Appendix, we provide a complementary investigation of accounting variables. Using
quarterly data from Compustat, we compare trends between the UV and UL ﬁrms – over the twelve quarters
before and after the event date – in the following variables: assets, total liabilities/total assets (a measure of
leverage), plant, property and equipment, sales, the dividend ratio, Tobin’s average Q, proﬁt margins, and
44On average, there are 414/(39/1.25) ≈ 13 UV ﬁrms at any given point in time within the 15-month post-election window. One
way to benchmark this number is to suppose the true coefﬁcient of variation in the 15 month returns were 0.215 (the estimate from
Panel A of Table IV). This would imply one would need (1.65/0.215)2 ≈ 59 independent UV events at any given point in time to
secure positive returns 95 percent of the time. A probability of 0.99 would require (2.33/0.215)2 ≈ 117 independent UV events at
the same time.
45Shleifer and Vishny (1997) point out that, in theory, one could still argue that a large number of tiny arbitrageurs taking a small
position in this particular mispricing in a portfolio of arbitrage strategies across various markets could eliminate the union mispricing
(as well as all the other anomalies). But they argue that in the real world, arbitrage involves few specialized investors taking large
positions.
46Barberis and Thaler (2003) also consider the case where the substitute security does eliminate fundamental risk so that only
noise trader risk remains. They point out that, even in that case, mispricing can persist if, in addition, arbitrageurs effectively have
short horizons, which, as argued in Shleifer and Vishny (1997), is the case in the real world where specialized arbitrageurs are
evaluated by investors not by their strategy but by their (short-run) returns.
47While Fama (1998) questions the robustness of some of these ﬁndings, he acknowledges that the slow post-earnings announcement
drift “has survived robustness checks, including extension to more recent data.”
19
the returns on assets. We ﬁnd some evidence that UV ﬁrms exhibit a downward break in trend (relative to
UL ﬁrms) in total assets, shareholder equity and sales. On the other hand, we ﬁnd evidence that returns on
assets and proﬁt margins remain stable, though these estimates are fairly imprecise. While these patterns are
generally consistent with our event study analysis, we view the evidence as suggestive, since the data are at
lower frequencies and more noisy. A more detailed discussion is in the Online Appendix.
III.C Heterogeneous Impacts of Unionization
In view of the ﬁndings summarized in the preceding discussion, a natural question comes to mind: how can
these large effects be consistent with the substantially smaller effects found in DiNardo and Lee (2004)?
This sections aims at providing a partial answer to this question.
DiNardo and Lee (2004) exploits the “near-experiment” generated by secret ballot elections, comparing
establishments where unions became recognized by a close margin of the vote with workplaces where the
union barely lost; the analysis’ most precise estimates are those for wages: increases of 2 percent could be
statistically ruled out as far away as seven years after the election.48
There a number of reasons for the apparent divergence between those results and the analysis reported
here. For one, it may take a much longer period of time – perhaps a decade or more – for unions to establish
enough support within the workplace to have the required bargaining power to negotiate for substantially
higher wages. Secondly, unions impose other costs that are not measured by the LRD, such as the use of
seniority rules, work rules, grievance procedures, and other working conditions speciﬁed in union contracts.
In principle, our approach in this paper of examining the effect of stock market valuation addresses both of
these concerns: if the market correctly prices the ﬁrm, it should capture the sum of all costs imposed by
the union, and effects that might occur many years in the future should be capitalized into the stock market
valuation of the ﬁrm in the relative short-run.
A ﬁnal important limitation of the RD analysis is that by estimating a discontinuity in the relationship
between wages and the vote share at the 50 percent threshold, it can only estimate a weighted average
treatment effect, where the weights are proportional to the ex ante likelihood an election was predicted to be
“close.”49 That is, among the observed close elections, a disproportionately small number would have had the
fundamentals of strong union support. The RD design is fundamentally unable to provide a counterfactual
48Interestingly, the magnitudes are also in line with what was found on wages in Lalonde et al. (1996). Freeman and Kleiner
(1990b) also ﬁnd wage effects that are much smaller than those found in cross-sectional worker-level studies.
49For a detailed discussion of this interpretation, see Lee (2008).
20
for the set of elections where workers voted 90 percent in favor of unionization.
But because our counterfactual is what would have happened had there not been an election at all, we can
directly examine the heterogeneity in the effects of unionization at all points in the vote share distribution.
This analysis is possible because of the long-panel structure we have at our disposal.
We begin by relating the security-level cumulative abnormal return in the two years following the election
to the union vote share. Speciﬁcally, we are interested in the shape of E[CAR(0,24)i|vi], where vi denotes
the union vote share in election i. We graphically plot this function by: (1) averaging CAR(0,24)i over 20
equally-spaced vote share bins50 and (2) plotting the predicted values from the model E[CAR(0,24)i|vi]=
p(vi) + β1(vi > 0.5), where p(·) denotes a sixth-order polynomial and 1(vi > 0.5) is an indicator function
for whether the union vote share in a given election exceeded 50 percent. Figure VII presents estimates of
E[CAR(0,24)i|vi] using both of these approaches. (For reference, Online Appendix Figure III shows the
histogram of the union vote share variable.)
Figure VII shows clear evidence that the effect of a certiﬁcation election is heterogeneous, and that it
depends on the union vote share. As in the Dinardo and Lee study, there is no discernible discontinuity
in the E[CAR(0,24)i|vi] at the 50 percent union vote share threshold. In fact, the estimated discontinuity
is somewhat perverse: ﬁrms with close union victories experience elevated post-election cumulative returns
vis-a-vis ﬁrms with close union losses. On the other had, union victories with higher union vote shares
correspond to negative excess returns, and the negative impact of a union election win appears to become
markedly more pronounced when the union has a higher vote share. A greater than 60 percent union vote
share is associated with negative cumulative abnormal returns of 20 to 30 percent.
Firms with union losses also exhibit a downward sloping relationship between abnormal returns and vote
share. Much of the decline appears to occur at the largest vote shares, but there is also greater variability in
the predicted cumulative abnormal returns due to small sample sizes. Close union losses are associated with
marginally-signiﬁcant negative abnormal returns, though as we will show, these declines can be explained
by a small amount of pre-election trending in the abnormal returns.
We now turn to several robustness checks. In Figure VIII we overlay the predicted CAR in months 0
through 24 (shown in Figure VII) with the predicted CAR computed over event-months -24 to -4. The ﬁgure
shows that the gradient in CAR by vote share, seen for months 0 to 24, is not present for months -24 through
-4. This plot reassures us that the negative CAR observed for higher union vote shares is not a continuation
50See DiNardo and Lee (2004) for a description of the construction of these 20 equally spaced bins.
21
of a pre-event trend.
In Table V we conduct formal statistical inference. Using the same sample of 1,436 elections used to
construct Figure VII, in Column (1) we regress CAR(0,24) on a dummy for whether the union won the election.
Consistent with earlier analyses, we ﬁnd that union victories are associated with cumulative abnormal
returns that are 12.1 percentage points lower than ﬁrms with union losses (t-ratio = -3.5). In Column (2) we
add the union vote share as a covariate.51 The introduction of this variable alone is enough to change the sign
on the coefﬁcient of the union victory dummy, resulting in a union effect of 0.048 (t-ratio = 0.89). Adding
higher-order polynomial terms in the vote share (Column 3) only makes the estimated union victory coefﬁcient
more positive; the “regression discontinuity" estimate of a union victory is 8 percentage points, but
is statistically indistinguishable from 0. In Column (4) we examine whether the negative gradient between
CAR and the vote share differs among elections where the union won and lost. Speciﬁcally, we regress
CAR(0,24) on a union victory indicator, the vote share, and the vote share interacted with the win indicator.
The interaction term is statistically insigniﬁcant in all speciﬁcations. In Columns (5)-(8) we estimate the
same set of models using CAR(-24,-4) as the dependent variable. None of the patterns observed when using
CAR(0,24) as the dependent variable are evident here.
The larger market value changes associated with large-margin union victories may at ﬁrst seem surprising,
since one would expect that the likelihood of a successful organizing attempt would have been known
to be very high for these ﬁrms, and if the victory was almost a forgone conclusion, the impact of the union
victory would already be priced into the stock by the date of the election. If one were concerned with this
possibility, then one could accumulate the returns starting at a point well before the original petition. At
this earlier, pre-organizing-drive date – even if the probability of victory conditional on an organizing drive
occurring was quite high – the overall probability of a union organizing attempt in the ﬁrst place and a union
victory could be quite small. As it turns out, as Figures II and III show, it makes little difference whether
the abnormal returns are accumulated starting at 15 months prior to the election – which we believe is well
before an organizing drive would even begin (see Roomkin and Block (1981)) – or starting at month 0.
51Vote share is grouped into one of 20 equally spaced bins, ranging from 0 to 1. We transform this variable to avoid the “integer”
problem described in DiNardo and Lee (2004).
22
IV Interpretation and Policy Implications
Here we brieﬂy summarize the results of our analysis of what our empirical results might imply about the
potential effect of a policy that makes it easier for workers to unionize. A much more detailed discussion can
be found in the Online Appendix. An example of a policy shift that could potentially ease unionization is the
so-called Employee Free Choice Act, recently proposed legislation that is meant to amend the National Labor
Relations Act. Speciﬁcally, one of the provisions of the legislation would allow employees to authorize a
union via “card check”, a showing that the majority of the workers signed cards to authorize a union, without
having to win certiﬁcation via a secret-ballot election process.52 It is widely believed that the legislation,
supported by the AFL-CIO, would make it much easier for workers to unionize, if it were to become law.
How the dynamics of unionization might change under such a law is unknown, and indeed, if certiﬁcation
is based on a showing of union authorization cards, ﬁrms may expend more resources to oppose card drives.
Nevertheless, there are a few reasons – apart from the support that this proposal has received from the AFLCIO
– why one could expect the law to ease unionization. First, Riddell (2004) provides some empirical
evidence from British Columbia that unionization rates signiﬁcantly fell when the card-check procedure was
replaced with a system of U.S.-style elections, and then increased by the same amount, when card-check
was restored. Second, however differently the ﬁrms respond to such a new regime, it is clear that the number
of available actions the ﬁrm can take to oppose unionization would strictly decrease under the proposed
legislation (i.e. under current law, they can already expend resources to try to discourage a signature drive).
Currently, even though having signatures from 30% of the workplace is required at the petitioning stage,
unions do not usually attempt to unionize unless they have signatures from signiﬁcantly more than 50%, as
they anticipate a drop in support throughout the election campaign. Under the EFCA scenario there would
no longer be elections, but it is still true that we can view workers as deciding between two options (sign
card or not). If EFCA strictly eases the path of unionization as we have argued, then this is not unlike an
election with a lower vote threshold. Thus, in our simulations, we consider a ceteris paribus lowering of the
vote threshold for certiﬁcation.
As a thought experiment, consider lowering the threshold from 50 percent to say, 45 percent. One
conjecture is that such a policy change would only effect those ﬁrms with vote shares between 45 and 50
52Under the proposed EFCA, a successful “card check” would also guarantee a ﬁrst contract because failure to sign a contract
within 120 days would result in binding arbitration to establish one. Our stylized model and simulation focuses on the ease of
certiﬁcation and abstracts from this and other aspects of the legislation.
23
percent, and that the effect could be approximated by the RD estimate. The shortcoming of this conjecture is
that it assumes that unions, ﬁrms, and workers do not respond to the increased ease of unionization. As we
noted in the introduction, Friedman (1950) suggested that unions – aware of potential employment effects
– might temper wage demands when seeking the support of their workers. In a representation election, this
might mean moderating wage expectations to increase their chance of winning. With these forces at work, an
exogenous increase in the probability of a union victory could very well lead unions to be more aggressive,
resulting in increased negative impacts on proﬁtability – not just for those ﬁrms near the 50 percent threshold,
but also for those where the union won by a wider margin. Exogenously easing the unionization process
might also affect the outcome for ﬁrms that eventually do not unionize, through union threat. Thus, to
make quantitative predictions regarding the impacts of making unionization easier – predictions that both
use the magnitudes we estimate, and allow for behavioral responses to a change in policy – it is necessary to
adopt assumptions about the behavior of unions and ﬁrms and how proﬁtability is affected by changes in the
probability of unionization.
We consider a “median voter”-type model of endogenous union determination. In the model, the union
and ﬁrm propose a wage level (which in our model has a one-to-one correspondence with proﬁt levels),
and voters (the workers), recognizing that wages can be both too low or “too high” (if it poses too large
a risk of consequent job loss), vote on the two choices in the election. Both the union and the ﬁrm face
trade-offs: the union (ﬁrm) would beneﬁt from higher (lower) wages, but proposing those wages loses votes
among those workers who have more moderate preferences. We consider a Nash Equilibrium whereby the
union (management) maximizes expected utility (or proﬁt) with the correct anticipation of the management’s
(union’s) proposal. The model is very similar to the model of ﬁnal offer arbitration in Farber (1978), which
involves two bargaining parties (here, the union and management) and an arbitrator (here, the median voter)
whose notion of the fair award is uncertain from the perspective of the two parties (and is represented by a
known probability distribution). The model formalizes the notion that the effects can be much larger for cases
in which the union wins by a large margin, because in those cases, workers have much stronger preferences
for wages.53
In the Online Appendix, we develop and discuss this model, showing that the most parsimonious speciﬁcation
of this kind of model can be given by 6 parameters: the distribution of “median voter” preferences
53Thus, in this framework, the ultimate monopoly power of the union stems from the location of the underlying exogenouslydetermined
preferences of the voters.
24
across workplaces (µ,λµ), the degree of uncertainty (from the ﬁrms’ perspective) of the precise position of
the median voter (λε), the degree of heterogeneity in workers’ preferences within a ﬁrm (σ), the implicit
limit on how low the ﬁrms’ wages can be (π), and the union’s “ideal” proﬁt level (or, equivalently, ideal
wage level) (c).We then calibrate it by choosing parameters such that the model produces both an equilibrium
vote share distribution and event-study estimates that most closely match that which we observe in the
data, speciﬁcally the patterns shown in Figure VII and the shape of the distribution of vote shares (shown
in Appendix Figure III). Viewed through the lens of this electoral competition model, the results imply that
union and the ﬁrm “offers” (and hence realized outcomes) are generally more “moderate” than the positions
of the median voters. Overall, the electoral competition reduces the variability of outcomes across workplaces,
and in the relatively small share of elections where the union effect is large, workers’ preferences are
more extreme than what the union offers.54
We then use these parameters to simulate the effect of easing unionization by changing the vote share
threshold from the current 50 percent threshold to lower thresholds (33, 25, and 10 percent). Our analysis
predicts that a change in the threshold from 50 to 33 percent would approximately double the victory rate
from 33 percent to 70 percent, and reduce the market value of all ﬁrms by 4.3 percent. In the simulation,
most of this effect is driven by marginal ﬁrms that are newly unionized due to the shift, but there is a modest
negative effect on the inframarginal ﬁrms that would be unionized under either scenario. A more dramatic
policy that would lower the threshold to 10 percent – which would imply virtually a victory rate of about 99
percent would lead to an 11 percent decline in the market value across all ﬁrms. We note that the magnitudes
might be somewhat overstated because the change in market values for the marginal group are reasonably
approximated by the discontinuity apparent in the simulated data, which are somewhat larger than the point
estimates we obtain from our RD analysis in Section III.C.
V Conclusion
The economic effects of unions on the labor market and the economy have been a longstanding area of
interest for economists. The literature has considered the impact of unions on wages, their potential role as
54Our model is a simple representation of the trade-offs facing the workers, the employer, and the union when they are acting
strategically. There are alternative models that one could consider that could explain certain features of the data. For example, it
is possible that these parties are not acting strategically and the pattern of estimates are explained by unions needing substantial
support from workers in order to be effective in negotiating a contract. While ﬁrms in the model can resist unionization, it is through
their ability to make offers to the workers. An alternative and worthwhile extension would be to allow ﬁrms to mount campaigns
against union drives, as in Freeman and Kleiner (1990a).
25
monopolies, their role in work stoppages, their effect on the aggregate economy, as well as the question of
how they can even exist and survive in a competitive labor market. To even partially address many of these
questions, we must ﬁrst understand how unions affect ﬁrms.
We began by asking whether the case of National Linen Services was the rule or the exception. In one
respect, it is the rule. We have shown that among publicly traded ﬁrms where the workforce attempting to
organize is not too small, new unionization is associated with a reduction in the ﬁrm’s market value, in a
way that parallels the experience of NLS. Like the NLS case, the stock market reaction to union victories is
somewhat slow, as has been found in a number of other event-study contexts. This ﬁnding is robust to the
use of a variety of speciﬁcations and to the use of several different methodologies. The negative effects
of unionization on the equity value of ﬁrms appears fairly stable over time, showing no major differences
before or after 1984.
In another respect, however, the case of NLS is a clear exception. By two years after the union victory,
NLS stock had earned negative 75 percent abnormal returns. By contrast, for our sample we estimate abnormal
returns of about negative 10 percent, and our sample is somewhat representative of publicly-traded
ﬁrms at risk of unionization. Based on the market capitalization of these ﬁrms, this 10 percent equity loss
translates to a total loss of about $40,500 (in 1998 dollars) per newly unionized worker. Since this amount
represents a combination of a transfer to workers as well as lost proﬁt due to inefﬁciencies caused by the
union, one can view this magnitude as an upper bound on the redistributive effect or the efﬁciency effect.55
For example, if the true average union wage effect is 8 percent and if our back-of-the-envelope calculation
(that a $40,500 loss would translate to a pure transfer equivalent to a 10 percent wage premium) is correct,
then this would imply a 2 percent loss in terms of efﬁciency due to unions.
The large difference in magnitude between the case of NLS and the estimated average effects serves to
highlight the importance of heterogeneous effects, which we carefully document in our analysis. Using a
different sample from DiNardo and Lee (2004), we also ﬁnd RD estimates that imply unionization is largely
ineffective for ﬁrms where there is more moderate support for unions, at least to the extent that unions do not
affect a ﬁrm’s equity value. This ﬁnding can be reconciled with the ﬁndings from the event-study analysis
through the negative gradient in abnormal returns in relation to the union vote share.
Finally, we consider a voting model of endogenous union determination, and calibrate it with the mag-
55Treating this magnitude as an upper bound requires assuming that unions can only impose efﬁciency costs, and cannot lead to
increases in proﬁtability (after netting out compensation costs).
26
nitudes we ﬁnd in our empirical analysis to make a ﬁrst-cut prediction on the likely impact of policies that
increase the likelihood of unionization. Policy simulations show that easing the threshold necessary to gain
recognition would not lead to union threat effects (ﬁrms losing value by having to respond to the threat of
unionization), but would cause unions to use this increased voter slack to be more aggressive. While the
RD estimates reasonably approximate effects for small policy changes, the approximation leads to a increasingly
larger understatement of the effects of larger policy shifts. Our exercise suggests that a policy-induced
doubling of unionization would lead to a 4.3 percent decrease in the equity value of all ﬁrms at risk of
unionization.
27
References
Abowd, John M., “The Effect of Wage Bargains on the Stock Market Value of the Firm,” American Economic
Review, September 1989, 79 (4), 774–800.
Allen, Steven G., “The Effect of Unionism on Productivity in Privately and Publicly Owned Hospitals and
Nursing Homes,” Journal of Labor Research, 1986, 4, 59–68.
, “Unionization and Productivity in Ofﬁce Building and School Construction,” Industrial and Labor Relations
Review, 1986, 39, 187–201.
Ashenfelter, Orley and David Card, “Time Series Representations of Economic Variables and Alternative
Models of the Labour Market,” Review of Economic Studies, 1982, 49 (5), 761–81.
Asquith, Paul, “Merger bids, uncertainty, and stockholder returns,” Journal of Financial Economics, April
1983, 11, 51–83.
Ball, Ray and Philip Brown, “An Empirical Evaluation of Accounting Income Numbers,” Journal of Accounting
Research, Autumn 1968, 6 (2), 159–178.
Barber, Brad M. and William John D. Lyons, “Detecting long-run abnormal stock returns: The empirical
power and speciﬁcation of test statistics,” Journal of Financial Economics.Volume 43, Issue 3, March
1997, Pages 341-372, March 1997, 43 (3), 341–372.
Barberis, Nicholas and Richard Thaler, “A Survey of Behavioral Finance,” in George M. Constantinides,
Milton Harris, and Rene M. Stulz, eds., Handbook of the Economics of Finance, Vol. 1B, Amsterdam:
Elsevier Science B.V., 2003, chapter 18, pp. 1051–1121.
Basker, Emek, “The Causes and Consequences of Wal-Mart’s Growth,” Journal of Economic Perspectives,
2007, Summer, 177–198.
Becker, Brian E., “Concession Bargaining: The Impact on Shareholders’ Equity,” Industrial and Labor
Relations Review, 1987, 40, 268–279.
and Craig A. Olson, “The Impact of Strikes on Shareholder Equity,” Industrial and Labor Relations
Review, April 1986, 39 (3), 425–438.
Bernard, Victor L. and Jacob K. Thomas, “Post-Earnings-Announcement Drift: Delayed Price Response
or Risk Premium?,” Journal of Accounting Research, 1989, 27, 1–36.
Blanchﬂower, David G. and Alex Bryson, “What Effect Do Unions Have on Wages Now and would Freeman
and Medoff Be Surprised?,” in James T. Bennett and Bruce E. Kaufman, eds., What Do Unions Do?
A twenty-Yr. Perspective, Transaction Publishers, 2007.
Brav, Alon, “Inference in Long-Horizon Event Studies: A Bayesian Approach with Application to Initial
Public Offerings,” The Journal of Finance, October 2000, 55 (5), 1979–2016.
Bronars, Stephen G. and Donald R. Deere, “Union Representation Elections and Firm Proﬁtability,” Industrial
Relations, Winter 1990, 29, 15–37.
Cameron, Colin A., Jonah B. Gelbach, and Douglas L. Miller, “Robust Inference with Multi-Way Clustering,”
Technical Working Paper T327, National Bureau of Economic Research September 2006.
28
Clark, Kim B., “The Impact of Unionization on Productivity: A Case Study,” Industrial and Labor Relations
Review, 1980, 33, 451–469.
, “Unionization and Productivity: Micro-Econometric Evidence,” Quarterly Journal of Economics, 1980,
95, 613–639.
, “Unionization and Firm Performance: The Impact on Proﬁts, Growth, and Productivity,” American
Economic Review, December 1984, 74 (5), 893–919.
Cooke, William N., “The Failure to Negotiate First Contracts: Determinants and Policy Implications,”
Industrial and Labor Relations Review, January 1985, 38 (2), 163–178.
Cusatis, Patrick J., James A. Miles, and J. Randall Woolridge, “Restructuring through spinoffs: The
stock market evidence,” Journal of Financial Economics, June 1993, 33 (3), 293–311.
Dellavigna, Stefano and Joshua Pollet, “Demographics and Industry Returns,” American Economic Review,
December 2007, 97 (5), 1167–1702.
Dharan, Bala G. and David L. Ikenberry, “The Long-Run Negative Drift of Post-Listing Stock Returns,”
The Journal of Finance, December 1995, 50 (5), 1547–1574.
DiNardo, John and David S. Lee, “Economic Impacts of New Unionization on Private Sector Employers:
1984-2001,” Quarterly Journal of Economics, 2004, 119, 1383–1442.
and Kevin F. Hallock, “When Unions "Mattered": The Impact of Strikes on Financial Markets, 1925-
1937,” Industrial and Labor Relations Review, 2002, 55.
Ellwood, David T. and Glenn Fine, “The Impact of Right-to-Work Laws on Union Organizing,” The Journal
of Political Economy, 1987, 95(2), 250–273.
Fama, Eugene, “Market efﬁciency, long-term returns, and behavioral ﬁnance,” Journal of Financial Economics,
September 1998, 49 (3), 283–306.
Farber, Henry, “Individual Preferences and Union Wage Determination: The Case of the United Mine
Workers,” Journal of Political Economy, 1978, 86, 923–942.
Farber, Henry S., “Right-to-Work Laws and the Extent of Unionization,” Journal of Labor Economics, July
1984, 2 (3), 319–352.
, “Union Membership in the United States: The Divergence between the Public and Private Sectors,”
Industrial Relations Section Working Paper 503, Princeton University September 2005.
and Bruce Western, “Accounting for the Decline of Unions in the Private Sector, 1973-1998,” Journal
of Labor Research, 2001, 22, 459–486.
Featherstone, Liza, “Will Labor Take the Wal-Mart Challenge?,” The Nation, 2004, June.
Freeman, Richard B. and James L. Medoff, What Do Unions Do?, New York: Basic Books, 1984.
and Morris M. Kleiner, “Employer Behavior in the Face of Union Organizing Drives,” Industrial and
Labor Relations Review, April 1990, 43 (4), 351–365.
and , “The Impact of New Unionization on Wages and Working Conditions,” Journal of Labor Economics,
January 1990, 8 (1), S8–S25.
29
Friedman, Milton, “Some Comments on the Signiﬁcance of Labor Unions for Economic Policy,” in
David McCord Wright, ed., The Impact of the Union: Eight Economic Theorists Evaluate the Labor
Union Movement, New York: Harcourt, Brace and Company, 1950. Institute on the Structure of the Labor
Market, American University, Washington D.C.
Heckman, James and Joseph Hotz, “Alternative Methods for Evaluating the Impact of Training Programs,”
Journal of the American Statistical Association, 1989.
Hirsch, Barry T., Labor Unions and the Economic Performance of U.S. Firms, Kalamazoo: Upjohn Institute
for Employment Research, 1991.
, “Union Coverage and Proﬁltability among U.S. Firms,” Review of Economics and Statistics, 1991, 73,
69–77.
, “Reconsidering union wage effects: Surveying new evidence on an old topic,” Journal of Labor Research,
June 2004, 25 (2), 233–266.
, “What Do Unions Do for Economic Performance?,” in James T. Bennett and Bruce E. Kaufman, eds.,
What Do Unions Do? A Twenty-Yr. Perspective, Transaction Publishers, 2007.
and David A. Macpherson, Union Membership and Earnings Data Book, Bureau of National Affairs,
Inc., 2008.
Holmes, Thomas, “Geographic Spillover of Unionism,” April 2006.
Hong, Harrison, Terence Lim, and Jeremy Stein, “Bad News Travels Slowly: Size, Analyst Coverage,
and the Proﬁtability of Momentum Strategies,” The Journal of Finance, 2000, 55 (1), 265–295.
, Terrence Lim, and Jeremy C. Stein, “Bad News Travels Slowly: Size, Analyst Coverage, and the
Proﬁtability of Momentum Strategies,” Journal of Finance, 2000, 55 (1), 2650295.
Huth, William L. and Don N. MacDonald, “Equity Market Response to union Decertiﬁcation Petitions
and Elections,” Journal of Labor Research, Spring 1990, 11 (2), 193–201.
Ikenberry, David L., Graeme Rankine, and Earl K. Stice, “What Do Stock Splits Really Signal?,” The
Journal of Financial and Quantitative Analysis, September 1996, 31 (3), 357–375.
Jaffe, Jeffrey, “Special Information and Insider Trading,” The Journal of Business, July 1974, 47 (3), 410–
428.
Jegadeesh, Narasimhan and Sheridan Titman, “Returns to Buying Winners and Selling Losers: Implications
for Stock Market Efﬁciency,” The Journal of Finance, March 1993, 48 (1), 65–91.
Lalonde, Robert J., G. Marschke, and Kenneth Troske, “Using Longitudinal Data on Establishments to
Analyze the Effects of Union Organizing Campaigns in the United States,” Annales D’Economie et de
Statistique, 1996, 41/42, 155–185.
Lee, David S., “Randomized Experiments from Non-random Selection in U.S. House Elections,” Journal of
Econometrics, Forthcoming 2008, 142(2), 675–697.
Long, J. Bradford De, Andrei Shleifer, Lawrence H. Summers, and Robert J. Waldmann, “Noise Trader
Risk in Financial Markets,” Journal of Political Economy, August 1990, 98 (4), 703–738.
30
Loughran, Tim and Jay R. Ritter, “The New Issues Puzzles,” The Journal of Finance, March 1995, 50 (1),
23–51.
Lyon, John D., Brad M. Barber, and Chih-Ling Tsai, “Improved Methods for Tests of Long-Run Abnormal
Stock Returns,” The Journal of Finance, February 1999, 54 (1), 165–201.
Mandelker, Gershon, “Risk and return: The case of merging ﬁrms,” Journal of Financial Economics,
December 1974, 1 (4), 303–335.
Michaely, Roni, Richard H. Thaler, and Kent L. Womack, “Price Reactions to Dividend Initiations and
Omissions: Overreaction or Drift?,” The Journal of Finance, June 1995, 50 (2), 573–608.
Mitchell, Mark L. and Erik Stafford, “Managerial Decisions and Long-Term Stock Price Performance,”
The Journal of Business, July 2000, 73 (3), 287–329.
Mitchell, Merwin A. and Joe A. Stone, “Union Effects on Productivity: Evidence from Western Sawmills,”
Industrial and Labor Relations Review, 1992, 46, 135–145.
Moore, William J. and Robert J. Newman, “The Effects of Right-to-Work Laws: A Review of the Literature,”
Industrial and Labor Relations Review, July 1985, 38 (4), 571–585.
Neumann, George R., “The Predictability of Strikes: Evidence from the Stock Market,” Industrial and
Labor Relations Review, July 1990, 33 (4), 535–535.
Olson, Craig A. and Brian E. Becker, “The Effects of the NRLA on Stockholder Wealth in the 1930s,”
Industrial and Labor Relations Review, October 1990, 44 (1), 116–129.
Riddell, Chris, “Union Certiﬁcation Success under Voting versus Card-Check Procedures: Evidence from
British Columbia, 1978-1998,” Industrial and Labor Relations Review, 2004, 57, 493–517.
Roomkin, Myron and Richard Block, “Case Processing Time and the Outcome of Representation Elections:
Some Empirical Evidence,” Univeristy of Illinois Law Review, 1981, 1, 75–97.
Ruback, Richard and Martin Zimmerman, “Unionization and Proﬁtability: Evidence from the Capital
Market,” Journal of Political Economy, 1984, 92, 1134–1157.
Shleifer, Andrei and Robert W. Vishny, “The Limits of Arbitrage,” The Journal of Finance, 1997, 52 (1),
35–55.
U.S. Census Bureau, “Statistical Abstract of the United States: 2011 (130th Edition),” Washington, DC.
<http://www.census.gov/statab/www/> 2010.
Visser, Jelle, “Union Membership Statistics in 24 Countries,” Monthly Labor Review, January 2006, 129 (1),
38–49.
Wessels, Walter J., “Do unionized ﬁrms hire better workers?,” Economic Inquiry, 1994, 32(4), 616–29.
31
Data Appendix
This Appendix describes how we match establishments in the NLRB data to ﬁrms in the CRSP database.
When matching we looked for similarities in the name listed in the NLRB election ﬁle to names that were
ever present in the CRSP ﬁles. To this end, we created two data sets: one containing the company names in
the NLRB election ﬁle and the other containing every company name that has ever appeared in the CRSP
database.56 This second data set will be hereafter referred to as the “master names ﬁle." In addition to the
company names, the master names ﬁle also contains a unique company id, the “PERMNO”, which allows
for further matching to the CRSP and Compustat databases.
There are 195,889 certiﬁcation elections in the NLRB data set that could potentially be matched to
companies in the master names ﬁle. Because the matching process is tedious, and must almost entirely
be done manually, we excluded any election with less than 100 voters. This resulted in 24,709 ﬁrms in
the certiﬁcation election ﬁle that potentially matched ﬁrms in the master list of CRSP company names.57
These elections are comprised of 61 percent of all workers eligible to vote in NLRB certiﬁcation elections.
Using this smaller subset, ﬁrms in the election ﬁle were compared to ﬁrms in the master CRSP ﬁle using the
matching algorithm employed by DiNardo and Lee (2004), which makes use of the SAS SPEDIS function.
The algorithm matches company names in the NLRB ﬁle to company names in the master names ﬁle based
on a so-called “spelling distance,” which considers those comparisons with a spelling distance below a predetermined
threshold as candidate matches.58 The algorithm may match a company in the election ﬁle to
more than one company name in the CRSP ﬁle. In these cases we selected the lowest spelling distance as
the candidate match. If there was a tie in spelling distance between two candidate comparisons, we selected
one match at random.
Because we matched ﬁrms on names only, manual inspection of the matches revealed that our automated
procedure resulted in many matches that were obviously incorrect. Therefore, research assistants reviewed
56Many companies have multiple names.
57Because a ﬁrm can have multiple elections, this number includes multiple cases of the same ﬁrms. There are 18,344 unique
ﬁrm spellings, though there are fewer unique ﬁrm names because of misspellings and abbreviations.
58We refer the reader to ? for further details on this algorithm. That study relied heavily on the establishment’s street address,
which is unavailable here. Therefore, the spelling distance threshold was quite speciﬁc to that application. As a ﬁrst pass, we
modiﬁed the program to match only on ﬁrm name, and discovered that in this application, that same threshold led to “too many”
matches. As we describe below, we therefore augmented the process with a manual review.
32
every match and dropped those where they judged the two ﬁrm names as different companies.59,60 We then
collected all of the unmatched companies in the election ﬁle, from the initial set of 24,709, and attempted to
locate each one in Dun and Bradstreet’s Million Dollar Directory and the Lexis/Nexis’ Directory of Corporate
Afﬁliations for the year of election. This step identiﬁed subsidiaries of publicly traded parent companies,
and allowed us to spot companies that were dropped erroneously in the previous step.
We ultimately matched 7,693 elections from the NLRB election ﬁle to companies in the CRSP master
ﬁle. In 1,579 cases, the ﬁrm in the CRSP ﬁle was not publicly traded at the time of the election. After
excluding the private ﬁrms, our ﬁnal sample contained 6,114 elections, consisting of 20 percent of all workers
eligible to vote in NLRB elections.
To determine whether the matches appeared reasonable, we compared the reported two-digit SIC industry
code and the state of the establishment from the election ﬁle to the corresponding variables in the CRSP and
Compustat ﬁles, for industry and state respectively. Because companies are diversiﬁed, the main SIC code
for a company in the CRSP database need not be the same as the SIC code for a particular establishment in
the NLRB election ﬁle. Similarly, an establishment may not be located in the same state as the company’s
headquarters. However, the comparisons are reassuring: the two digit SIC codes in the two data sets are the
same for 50 percent of the matches, while 40 percent of the matches show the same state. For reference,
if we randomly pair companies from the ﬁnal NLRB data set to companies in the master names ﬁle that
were never matched to the NLRB data through our procedure, the corresponding match rate is 5 percent for
industry and 4 percent for state.
59For example, the algorithm determined that any company in the election ﬁle with the word “American” as part of its name was
a sufﬁciently good match for the company “American Enterprises” in the CRSP ﬁle, if a better match did not exist. Therefore, a disparate
set of companies like “American Laundry,”“American Envelope,”and “Pan American Screws” were all matched to “American
Enterprise.”All of these matches were dropped by our research assistants.
60Because there was an element of judgment, these exclusions were recorded in a log ﬁle for replication purposes.
33
Figure I
Cumulative stock market returns surrounding National Linen Service’s 1999 representation
election
Figure II
Average cumulative returns of union victory firms and of the size-matched reference portfolio, by
month relative to NLRB case closure
Note: Union victory firms consist of publicly traded companies holding representation elections where at least 5% of the
company’s workforce voted, and where the union won. Each point is the average cumulative return up to the month relative to
case closure, beginning 24 months prior to case closure. Each firm in the sample is associated with a benchmark portfolio
matched on size. The benchmark series corresponds to the average cumulative return of these size-matched reference portfolios.
Returns are expressed net of the risk-free rate.
Figure III
Average cumulative abnormal return of union victory firms, by month relative to NLRB case
closure
Notes: This figure shows the difference in the average cumulative return of union victory firms and the size-matched reference
portfolio, as shown in Figure 2. It corresponds to the average cumulative abnormal return computed beginning 24 months prior
to case closure. The dashed lines represent the 95% confidence intervals, which are computed using standard errors clustered on
elections and calendar months. We use the formula in Cameron, Gelbach, and Miller (2006) to compute standard errors with
multi-way clustering.
Figure IV
Average cumulative returns of union loss firms and of the size-matched reference portfolio, by
month relative to NLRB case closure
Note: Union loss firms consist of publicly traded companies holding representation elections where at least 5% of the company’s
workforce voted, and where the union lost. Each point is the average cumulative return up to the month relative to case closure,
beginning 24 months prior to case closure. Each firm in the sample is associated with a benchmark portfolio matched on size.
The benchmark series corresponds to the average cumulative return of these size-matched reference portfolios. Returns are
expressed net of the risk-free rate.
Figure V
Average cumulative abnormal returns of union loss firms, by month relative to case closure
Notes: This figure shows the difference in the average cumulative return of the union loss portfolio and the size-matched
reference portfolio, shown in Figure 4. It corresponds to the average cumulative abnormal return computed beginning 24 months
prior to case closure. The dashed lines represent the 95% confidence intervals, which are computed using standard errors
clustered on elections and calendar months. We use the formula in Cameron, Gelbach, and Miller (2006) to compute standard
errors with multi-way clustering.
Figure VI
Average cumulative abnormal return, by analyst coverage
Note: A company is considered to have analyst coverage if it appears in the I/B/E/S dataset in the year of the
election. The sample is limited to elections occurring in years where I/B/E/S data were available, between 1976 and
1999
Figure VII
Cumulative abnormal returns in the two years after NLRB closes election, by relation to vote share
Note: Abnormal returns are the simple difference in the security’s return and the size-matched benchmark portfolio
in the same month. Cumulative abnormal returns are the sum of the abnormal returns over a two year period
beginning in the month of case closure. Predicted values are calculated using a sixth-order polynomial, and an
indicator for whether the union won. Dashed lines are the 95% confidence interval. Dots are the average
cumulative excess return in 20 equally spaced bins. See Section 4.4 for further details on the construction of this
figure.
E[CAR(0,24)|Union	  vote	  share]	  
Figure VIII
Cumulative abnormal returns in the pre- and post-event periods, by relation to vote share
Notes: Predicted values are calculated using a sixth-order polynomial and an indicator for whether the union won.
The solid line corresponds to the predicted cumulative excess return in the two years following case closure,
conditional on union vote share. The dashed line corresponds to the predicted cumulative abnormal return
calculated starting 24 months prior to the election through four months prior to case closure, conditional on union
vote share. See Section IV.D for further details on the construction of this figure.
TABLE I
Summary Statistics, Matched NLRB-CRSP Data
At least 5% of workforce voting Less than 5% of workforce voting
Union victory
(UV firms)
Union loss
(UL firms)
Union victory
(UV firms)
Union loss
(UL firms)
Number of elections 414 1022 1163 2682
Vote share for union 0.62 0.35 0.64 0.35
[0.11] [0.10] [0.13] [0.10]
Number of voters 449.1 454.2 276.5 297.6
[534.9] [558.5] [263.4] [301.6]
Number eligible 496.0 494.0 286.4 317.9
[649.3] [638.9] [286.1] [330.4]
Fraction of employees voting 0.21 0.23 0.01 0.01
[0.21] [0.21] [0.01] [0.01]
Year of election 1975.2 1976.9 1974.9 1976.6
[9.17] [9.11] [9.24] [9.42]
Fraction in Manufacturing 0.78 0.75 0.79 0.81
Number of employees 3813.3 3430.8 68468.6 75284.6
[5377.5] [5195.4] [134336.5] [123610]
Market Value (CRSP) 353.8 330.9 4734.1 6350
[880.3] [783.8] [10,547] [13,660]
Fraction of stocks delisted 0.10 0.08 0.049 0.028
Notes: Summary statistics are based on the NLRB election and CRSP data. Standard deviations are in brackets.
Market value is in millions of dollars. Fraction of stocks delisted is computed as the fraction of stocks with a
non-missing delisting return in a two year window surrounding the NLRB case closure month.
TABLE II
Estimates of post-election cumulative abnormal returns
(1) (2) (3)
Size-matched
benchmark
Size × industry-
matched
benchmark
Broad-market
benchmark
Panel A: Union Victory
ACAR(0,24) -0.092 -0.096 -0.103
(0.028) (0.028) (0.029)
ACAR (-24,-4) -0.010 -0.009 -0.010
(0.020) (0.020) (0.020)
Panel B: Union Loss
ACAR (0,24) 0.029 0.020 0.016
(0.021) (0.020) (0.022)
ACAR (-24,-4) 0.034 0.004 -0.009
(0.022) (0.014) (0.014)
Notes: ACAR(X,Y) denotes the average cumulative abnormal return from month X to month
Y relative to the NLRB case closure month. There are 414 elections in the sample in Panel
A, and 1022 elections in Panel B. See Section III.B for details on the construction of the
benchmark portfolios and estimation.
TABLE III
Relating post-event cumulative abnormal returns to the share of the
workforce in the bargaining unit
(1)
Union victory
ACAR(0,24)
(2)
Union loss
ACAR(0,24)
Constant
0.03
(0.01)
0.03
(0.01)
Share of workforce in
bargaining unit
-0.31
(0.08)
0.06
(0.05)
Observations 1577 3704
Note: Sample includes all NLRB elections that we matched to publicly traded
firms. See note to Figure II for details on how ACAR(0,24) was constructed.
TABLE IV
Evaluation of Arbitrage Strategies
Arbitrage Strategy Buy and Hold Benchmark
Event Month
Coefficient of
Variation Percent Positive
Coefficient of
Variation Percent Positive
Panel A: Individual Investor Strategy
5 0.211 0.606 0.398 0.664
10 0.226 0.633 0.629 0.737
15 0.215 0.647 0.737 0.773
Panel B: Diversified Strategy
5 0.209 0.595 0.426 0.668
10 0.313 0.620 0.603 0.718
15 0.377 0.678 0.746 0.775
24 0.427 0.695 0.958 0.844
36 0.497 0.674 1.113 0.879
60 0.604 0.780 1.231 0.941
120 0.867 0.802 1.382 1.000
360 2.829 1.000 4.458 1.000
Notes: Panel A provides the coefficient of variation of the returns, and the percent of the time the returns are
positive for 414 investors. In the first two columns, each investor follows a zero-investment portfolio, holding a
short position in the UV firm and a long position in its benchmark portfolio. In the second two columns, the
CRSP stock index is held instead. Panel B provides the same statistics for a single diversified investor (as
described in the text) over all possible starting months within the sample period.
TABLE V
Cumulative abnormal returns in relation to vote share
CAR(0,24): CAR(-24,-4):
(1) (2) (3) (4) (5) (6) (7) (8)
Constant 0.029 -0.065 -0.075 -0.064 0.0003 -0.018 -0.029 -0.035
(0.021) (0.030) (0.039) (0.035) (0.016) (0.025) (0.033) (0.030)
Union
victory -0.121 0.048 0.080 0.049 0.003 0.037 0.021 0.027
(0.035) (0.054) (0.066) (0.053) (0.028) (0.046) (0.057) (0.047)
Union
victory -0.016 0.332
×vote share (0.321) (0.255)
vote share -0.616 -0.610 -0.123 -0.235
(0.160) (0.207) (0.126) (0.162)
p(vote share) X X
Observations 1436 1436 1436 1436 1436 1436 1436 1436
Note: Robust standard errors are in parentheses. The sample consists of all elections where at least
5% of the workforce voted. The variable “vote share” denotes the union vote share, minus 0.5.
Following Dinardo and Lee (2004), the vote share is aggregated to 20 discrete bins. The dependent
variable is the cumulative abnormal return from months 0 to 24 relative to case closure (columns 1-
4), and the cumulative abnormal return from -24 through -4 months relative to case closure (columns
5-8). The term p(vote share) denotes a fourth-order polynomial in the union vote share.