Page 1

I

Published by Blackwell Publishing, Inc., 350 Main Street, Malden, MA 02148, USA, and 9600 Garsington

Road, Oxford, OX4 2DQ, UK.

R

, Vol. 42, No. 4 (October 2003). © 2003 Regents of the University of California

650

What Do Unions Do to Productivity?

A Meta-Analysis

CHRISTOS DOUCOULIAGOS and PATRICE LAROCHE*

The impact of unions on productivity is explored using meta-analysis and

meta-regression analysis. It is shown that most of the variation in published

results is due to specification differences between studies. After controlling for

differences between studies, a negative association between unions and produc-

tivity is established for the United Kingdom, whereas a positive association is

established for the United States in general and for U.S. manufacturing.

T

HE

RELATIONSHIP

BETWEEN

UNIONS

AND

PRODUCTIVITY

has

attracted considerable attention from scholars in industrial relations and

economics, as well as from policymakers, unions, and business in general.

Despite voluminous theoretical literature, controversy continues regarding

the impact of unions on productivity, as well as on other aspects of busi-

ness, such as employment, research and development (R&D), profitability,

and investment. In traditional economic analysis, unions are said to dis-

tort labor market outcomes through, for example, legal and custom-driven

restrictions on relative wages, the imposition of employment restrictions,

and protection against layoffs. Unions are said also to be a contributing

factor to aggregate as well as sectoral unemployment and the associated

output losses. In contrast to these arguments, Freeman (1976) and Freeman

and Medoff (1984) argued that unions can raise productivity by providing

workers with a means of expressing discontent as an alternative to “exiting,”

by opening up communication channels between workers and management,

and by inducing managers to alter methods of production and to adopt

more efficient policies.

The controversy in the theoretical literature is matched by controversy

in the empirical literature. Empirical findings are divided between positive

and negative union-productivity effects, and many studies cannot reject the

hypothesis of a zero effect. Hence generalizations from the available evidence

* The authors’ affiliations are, respectively, School of Accounting, Economics and Finance, Deakin

University, Victoria, Australia, and Institut d’Administration des Entreprises, University of Nancy,

Nancy, France. E-mail:

douc@deakin.edu.au

and

patrice.laroche@univ-nancy2.fr.

Page 2

What Do Unions Do to Productivity? A Meta-Analysis

/ 651

are not obvious using traditional literature reviews. The aim of this article

is to make more lucid the relation between unions and productivity by using

meta-analysis and meta-regression analysis. Meta-analysis is now used

widely to identify and quantify patterns, draw inferences from the diversity

of results, and detect possible regularities in the association between unions

and productivity. Meta-analysis is used to implement a quantitative syn-

thesis of the available research and, where possible, to generalize from the

results derived from the numerous singular studies (Rosenthal 1984). Meta-

analysis is a scientifically valid statistical procedure that has been developed

to quantify associations drawn from an existing body of literature (Wolf

1986; Hunter and Schmidt 1990).

The meta-analysis presented in this article involves a comprehensive

survey and quantitative analysis of the published empirical literature.

Meta-regression analysis is used to examine the influence of methodologic

features and data differences on reported estimates of union-productivity

effects. Additionally, we explore the notion, prevalent in the empirical liter-

ature, of the existence of an invariant union-productivity effect.

While theoretical developments focus on efficient bargains—bargaining

over wages and work practices as well as bargaining over wages and

employment—empirical analysis of the net impact of unions on productiv-

ity remains of considerable interest. Hence meta-analysis of this empirical

literature is important. For example, policymakers have maintained their

interest in this area within the overall context of policy action and concerns

over labor market flexibility and labor market deregulation. Even though

the influence of unions has fallen and union membership has diminished in

most countries, industry in general continues to be concerned about the

impact of unionization, especially where productivity becomes important in

offsetting any adverse effect on cost competitiveness arising from the higher

wages of unionized labor. This has become even more imperative with the

rapid expansion of global markets. Furthermore, applied research in this

area continues, responding at least in part to these broader concerns. For

example, 37 studies estimating directly the impact of unions on productivity

were published in the 1990s compared with 31 published in the 1980s and

20 articles published in this area since 1995.

Previous Research and Reviews

There is now a sizable theoretical literature that explores both the hypo-

thesized costs and the benefits of trade unions. Examples include Addison

(1982, 1985), Addison and Barnett (1982), Freeman and Medoff (1984),

Page 3

652 /

C

D

P

L

Kuhn (1985), Hirsch and Addison (1986), Turnbull (1991), and Belman (1992).

Details of the theoretical arguments can be found in these and other sources.

Conceptually, unionism can booster or hamper labor productivity.

Because of the ambiguity over the net effect of unions, most of the existing

studies begin their empirical investigation without presupposing a specific

direction, leaving the conclusion to empirical findings. However, a common

conceptual framework serves as a starting point for empirical investigation.

This conceptual framework is the so-called two-faces view of unionism

(Freeman and Medoff 1984): the monopoly face and the collective voice/

institutional response face.

The

monopoly face

of unionism refers to a number of adverse wage and

nonwage effects. One of the most well established effects of unions is the

ability to increase wages above competitive levels (Lewis 1963). Another

dimension is unfavorable effects on R&D spending and tangible and intang-

ible investments. Union rent-seeking acts as a tax on the return on invest-

ment and limits innovative and investment activities (see Connolly, Hirsch,

and Hirschey 1986; Hirsch and Link 1987). These can have a detrimental

impact on the dynamic path of productivity.

Unions can have a direct negative impact on productivity by restricting

managerial discretion. For example, unions may force firms to adopt ineffi-

cient personnel hiring and firing practices. Legal restrictions against layoffs

and closed-shop arrangements have an impact on efficient factor usage

and hence productivity. Unions also may favor restrictive work practices,

curbing the pace of work, hours of work, and skill formation. They also may

obstruct the introduction of new technology (see McKersie and Klein 1983).

Productivity is affected also through strike activity. This arises through

working days lost, as well as noncooperative behavior that precedes or

follows strikes (see Flaherty 1987).

The other aspect of unions is the

collective voice and institutional response

face

(CV/IR) emphasized by Freeman and Medoff (1984). The CV/IR

model draws on the exit-voice dichotomy of Hirschman (1970). In this

framework: “voluntary quits become the labor market expression of exit,

and unions become the institution for the expression of (collective) voice”

(Turnbull 1991:137). By providing workers with a means of expressing dis-

content at the workplace, unions can reduce the extent to which quits and

absenteeism lead to a suboptimal degree of labor turnover. By presenting

unions as an alternative to resignation and apathy, Harvard scholars

delivered an argument in favor of union representation. This channel is

important because high labor turnover can reduce productivity in a work-

place through a direct loss of firm-specific training (see Addison and Barnett

1982; Freeman 1976).

Page 4

What Do Unions Do to Productivity? A Meta-Analysis

/ 653

According to Freeman and Medoff (1984), unions can enhance produc-

tivity by improving communication between workers and management. The

opening of communications channels between management and workers

can result in integrative rather than distributive bargaining (Dworkin and

Ahlburg 1985). Unions may provide additional information to a firm about

the preferences of employees, thus permitting the firm to choose a better

mix among working conditions, workplace rules, and wage levels. These can

result in a more satisfied, cooperative, and productive workforce.

In addition, unions may be responsible for a “shock effect.” Unions can

induce managers to alter methods of production and adopt more efficient

personnel policies (Slichter, Healy, and Livernash 1960). Union activities

also may improve worker morale and motivation. The potentially arbitrary

nature of decisions such as promotions or layoffs can be reduced by the

presence of unions. Consequently, the employee is more likely to see his or

her employer as fair. Leibenstein (1966) emphasized that one of the major

areas for improving X-efficiency in the firm is worker morale and motiva-

tion. Further, unions often stress seniority rules. There is a positive associ-

ation between productivity and experience. Seniority rules exclude a system

of subjective selection, and the seniority system emphasizes ability and

merit. It also may reduce conflict between seniority and efficiency (Rees

1962).

The CV/IR approach offers new insights into the role of unions in labor

productivity. This framework is based on a theory that “may be interpreted

as a hypothesis that internal organization matters and as an extension of

modern organization theory which abandons the standard textbook neo-

classical economic perception of the firm as a machine . . .” (Addison and

Barnett 1982:147).

The two faces of unions are not incompatible. Hirsch (1997:37) notes that

“The monopoly and collective voice faces of unionism operate side-by-side,

with the importance of each being very much determined by the legal and

economic environment in which unions and firms operate. For these rea-

sons, an assessment of union’s effects on economic performance hinges on

empirical evidence.”

Union productivity effects have been studied in a variety of industries,

including construction, manufacturing, mining, and services, as well as the

public sector. Most of the studies use U.S. or U.K. data, with many con-

flicting results reported. Unfortunately, the existing empirical studies do not

individually provide definitive answers on the relationship between unions

and productivity. These studies use disparate variables, methods, and

samples, and hence it is necessary to review the available studies and draw

inferences from them. Moreover, it is important to investigate the extent to

Page 5

654 /

C

D

P

L

which the differences between studies serve as potential explanations for the

disparity in the results across the studies. The differences between study

results may reflect actual differences in the relationship between unions and

productivity. The differences also could be due to the nature of the research

process.

There have been a number of reviews of this literature, some conducted

by leading experts. The conclusions drawn in the major reviews are broadly

similar. For example, in their review, Addison and Hirsch (1989:79) argued

that “Based on the extant evidence, we conclude that the average union

productivity effect is probably quite small and, indeed, is just as likely to be

negative as positive.” They note also that “. . . there is no compelling evi-

dence that, in general, the net effect of unions on productivity is positive or

negative” (1989:83). In his review, Kuhn (1998:1048) concluded that “A fair

summary of the industry studies is that most estimates are positive, with

the negative effects largely confined to industries and periods known for

their conflictual union-management relations, or to the public sector.” Sim-

ilar conclusions that the evidence supports neither a negative nor a positive

relationship are drawn by many other authors (see, for example, Wilson

1995). Preempting the need for meta-analysis, in the course of her review,

Booth (1995:223) noted that it is “necessary to have evidence on the

union effects from a number of different studies before drawing any firm

conclusions.”

The problem with qualitative reviews of any literature is that they may

suffer from what Stanley (2001) calls “casual methodological speculation.”

Since they are qualitative, they are based usually on opinion, and con-

clusions are drawn largely from subjective interpretation of the available

evidence, even when specialists conduct them. This makes qualitative reviews

prone to a greater degree of speculation than quantitative reviews. The

absence of statistical investigation of empirical results means that qualita-

tive reviews lack formal testing of a hypothesis. While qualitative reviews

assist with the subjective assessment of a hypothesis, it is only through a

quantitative review that a contentious hypothesis can be tested formally.

In contrast to existing reviews, in this article we present the first

tative

review of the union-productivity effects literature through meta-

analysis. Meta-analysis is used to “summarize, evaluate and analyze empirical

economic research” (Stanley 2001:131). It is well know that methodologic

and data differences have an impact on empirical estimates. The issue is how

to quantify that impact. Meta-analysis is based on a pronounced examina-

tion of differences in specification and datasets and is used to quantify the

impact these have on productivity effects. Meta-analysis helps to make sense

out of the substantial variation in union-productivity estimates. It is very

quanti-

Page 6

What Do Unions Do to Productivity? A Meta-Analysis

/ 655

difficult to evaluate the multidimensional nature of empirical investigations

using traditional literature reviews. For example, reviewers are forced to

assess the impact of specification differences without statistical tools that

enable them to identify the impact of specification differences after control-

ling for, say, data differences.

Meta-analysis should be seen as complementary to traditional reviews, a

way to analyze estimates and explain the variation of interstudy differences

(see Espey 1998). Despite differences in the review process, we show in this

article that the conclusions drawn from the existing reviews are correct with

respect to the entire pool of evidence. However, we can draw different con-

clusions about the direction of the productivity effects for important sub-

samples of this literature. In particular, it is possible to conclude that unions

have a negative impact on productivity in the United Kingdom and Japan

and that unions have a positive impact on manufacturing in the United

States. Importantly, in contrast to some of the qualitative reviews, the avail-

able evidence indicates that some of the productivity effects are of economic

significance.

Meta-Analysis and Meta-Regression Analysis Methodology

Meta-analysis was developed to facilitate a

thesis. There is now a burgeoning reference literature on meta-analysis [see, for

example, Cook et al.

(1992); Hedges and Olkin (1985); Hunter and Schmidt

(1990); Wolf (1986)]. Stanley (2001) offers a recent review of the growing

list of applications of meta-analysis in economics.

There are four steps in meta-analysis. Meta-analysis involves identifica-

tion and calculation of the association between variables of interest (known

as an

effect size

) by considering all the available relevant empirical literature.

Hence the first step in any meta-analysis is identification of the relevant

empirical literature. In the present study this is the published literature on

union-productivity effects. The second step involves derivation of effect

sizes from each study or calculation of effect sizes from information provided

in each study. Two effect sizes are used in this article, the partial correlation

coefficients between unions and productivity and the union-productivity

effects. The third step is calculation of summary statistics relating to the

effect sizes. The fourth step is moderator analysis—the identification of the

sources of variation between published effect sizes.

The most common approach known as

lation of summary statistics involving the associations of interest. The

key statistics of interest are the mean, the weighted mean, a measure of

quantitative

research syn-

meta-analysis

involves the calcu-

Page 7

656 /

C

D

P

L

homogeneity of research results, and confidence intervals constructed

around the mean. A separate branch known as

is used to uncover the sources of heterogeneity of research results. This

involves regressing study characteristics on the effect size derived from the

studies against a set of potential explanatory or moderator variables. Both

types of meta-analyses are presented in this article, with the emphasis on

meta-regression analysis.

meta-regression analysis

The Database.

database. We compiled all the published studies exploring the relationship

between unions and productivity. There are a number of unpublished

studies exploring this relationship, but these have not been included. Results

presented in unpublished material, such as manuscripts and working papers,

can change by the time they reach their published form, and hence in many

cases they may be less reliable than those found in published material.

An extensive computer search was conducted for studies written in Eng-

lish or French. A total of 73 statistically independent studies was compiled,

and these studies are used in the meta-analysis and meta-regression ana-

lysis.

In meta-analysis, studies are regarded as statistically independent if

different authors produce them or when they are by the same author but

different samples are used (see Hunter and Schmidt 1990). There are actu-

ally more than 73 published studies in this area. However, in some cases the

same authors have produced more than one published work using the same

dataset. These studies cannot be regarded as statistically independent. The

approach taken in this article was to average these non-statistically inde-

pendent studies and treat them as a single study. For example, Fitzroy and

Kraft (1987a, 1987b) use the same data, as do Guthrie (2001) and Guthrie,

Spell, and Nyamori

(2002).

From the point of view of meta-analysis, when two different authors use

the same dataset, both studies are regarded as statistically independent.

This is standard practice in meta-analytic reviews [see, for example, Espey,

Espey, and Shaw (1997) and Thiam, Brave-Ureta, and Rivas (2001)].

All the studies included in the meta-analysis provide direct measures of

the association between unions and productivity, with productivity as the

dependent variable and unionism as part of a set of explanatory variables.

The starting point for meta-analysis is compilation of the

1

2

1

Moreover, the quality of working papers varies. For example, those from the National Bureau of

Economic Research (NBER) are of a very high quality, and most of these have been published. How-

ever, there are working papers from other centers that have remained unpublished after many years and

in some cases after decades. Nevertheless, it is the case that most of the working papers have been

published and are included in the meta-analysis presented in this article.

2

A number of databases were searched, including EconLit, Proquest/ABI Inform, and EBSCO.

Page 8

What Do Unions Do to Productivity? A Meta-Analysis

/ 657

A number of empirical studies were excluded from the meta-analysis. These

include (1) the extensive body of literature exploring the impact of unions

on wages, (2) studies that explore the links between unions and employ-

ment, profitability and/or investment, (3) studies that explore the links

between unions and productivity but do not provide sufficient information

from which effect sizes could be calculated, (4) macroeconomic studies that

focus on the relationship between corporatist economic policies and eco-

nomic performance, (5) studies that explore the association between labor

relations climate and productivity through strike activities, grievances pro-

cedures, and quality of working life, (6) estimates of the union-productivity

growth

effect,

and (7) studies using probit models because they are not

comparable with the included studies. A full list of excluded studies is avail-

able from the authors.

3

Effect Sizes.

correlation coefficient, and from most of the studies we were able to collect

some information on the union-productivity effect.

measures of the association between the variables of interest (the effect sizes).

That is, the focus of the meta-analysis is the partial correlation between

unionization and productivity, as well as the productivity effects. These

measure the strength and, importantly, the direction of association between

unionization and productivity while holding other factors constant. The

techniques developed for traditional forms of meta-analysis are based on

measures such as correlations. Meta-regression analysis can be used for

both correlations and measures more favored by economists, such as elas-

ticities. One of the benefits of analyzing partial correlations is that it facili-

tates comparisons with other meta-analyses of workplace interventions and

performance. Examples where correlations are used include the meta-analysis

of job satisfaction and productivity (Miller and Monge 1986); absenteeism

and job performance (Bycio 1992); the impact of profit sharing, employee

share ownership, and employee participation in decision making (Doucouliagos

1995); and board of directors size and financial performance (Dalton et al. 1999).

From each of the published studies we calculated the partial

4

These are the preferred

3

Where studies reported results for both growth and levels, only the later was used.

None of the 73 studies reported partial correlation coefficients. However, they do report regression

coefficients, standard errors,

t

-statistics, or levels of statistical significance. The partial correlation coef-

ficients are calculated by using the

t

-statistics reported in the primary studies. Where

reported, they can be calculated from the reported levels of statistical significance or from the reported

regression coefficients and standard errors. The formula used to calculate partial correlations is:

where

t

is the

t-

statistic and

df

is degrees of freedom. Note that this will always produce a positive

number, so it is necessary to convert it to a negative number if the regression coefficient is negative (see

Greene 2000:Chap. 6).

4

t-

statistics are not

,

ttdf

22

/( )

+

Page 9

658 /

C

D

P

L

Most studies report more than one set of results. The approach adopted

in this article is to use only the results deemed by the study’s author or

authors as the preferred result. Hence we ignore any results undertaken for

sensitivity analysis, for exploration purposes, or just out of speculation.

Where studies report more than one valid regression coefficient, we aver-

aged these—in some cases the weighted average was used when the same

author in the same article used different sample sizes.

Meta-Regression Analysis

is to identify moderator variables and to explore the impact of specifica-

tion on the estimated union-productivity effect. Each of the 73 primary

studies used regression analysis to estimate union-productivity effects.

Meta-regression analysis is a regression analysis of the regression analysis

reported in the existing pool of studies. The published studies used a

number of different specifications, introduced different control variables,

and used different datasets covering different levels of aggregation, different

time periods, and different industries. Meta-regression analysis can be used

to detect whether these study characteristics are associated in any way with

the estimated study outcomes. This enables a

impact of differences in research design, methodology, data, and estimation

on reported study outcomes that is not possible in a traditional narrative

and qualitative literature review.

Meta-analysis also can be used to identify moderator variables. However,

the advantage of meta-regression analysis is its multivariate context. Meta-

regression analysis allows researchers to identify, for example, the associ-

ation between data aggregation (e.g., firm-versus industry-level data) and

estimated union-productivity effects while also controlling for other study

characteristics, such as the time span of data. Meta-regression analysis

offers a richer framework through which an existing body of empirical

literature can be reviewed.

The basic meta-regression equation takes the following form:

The principal use of meta-regression analysis

quantitative

assessment of the

Y

i

=

α

+

β

1

N

i

+

γ

1

X

i

1

+

…

+

γ

k

X

ik

+

δ

1

K

i

1

+

…

+

δ

n

K

in

+

ui

(1)

where Y = the partial correlation (or elasticity) derived from the ith study

α = a constant

Ni= sample size associated with the ith study

X = dummy variables representing characteristics associated with

the ith study

K = mean values of quantifiable variables, such as union density

ui = the disturbance term, with usual Gaussian error properties (see

Stanley and Jarell 1998)

Page 10

What Do Unions Do to Productivity? A Meta-Analysis

/ 659

Meta-Analysis Results

The 73 studies are presented in alphabetical order in Table 1 together

with the country to which the data relate, the sample size N used in each

study, the t-statistic (or the average t-statistic in cases where more than one

estimate is used per study), the partial correlation coefficient r, and the

associated union-productivity effect, as well as the publication outlet. The

union-productivity effects are presented in three separate ways. In column

6 we list δ, the establishment effect or the elasticity of productivity with

respect to union density. This is the preferred elasticity measure because it

measures the percentage change in labor productivity for an increase in

union density of 1 percent. As can be seen from Table 1, this elasticity is

not available from most of the studies. For most of the studies that did

report δ, we also present the mean union-productivity effect—that is, we

evaluate the impact of unions on productivity at sample means. Finally, we

also present the total productivity differential. This is available for most of

the studies and hence will be the central focus of the meta-regression ana-

lysis. This effect measures the impact of unions in the case of 100 percent

unionization. Studies are divided into those using density and those using

a dummy variable to denote union presence. The regression coefficients on

these are not comparable because the density studies measure δ, whereas

the dummy studies measure the impact on productivity arising from 100 per-

cent unionization. While it is true that 100 percent unionization is rare,

by evaluating the total productivity effect, we can compare most of the

studies.5 It can be seen from Table 1 that there is a wide range of results,

with both positive and negative findings. There is no apparent consistency

in the results. The partial correlations from 29 of the 73 studies are not

statistically significant. In 45 of the 73 studies a positive relationship was

found between unions and productivity, and the remaining 28 found a neg-

ative relationship. Of these, 26 found a positive and statistically significant

effect, whereas 18 found a negative and statistically significant effect. The

highest positive partial correlation is +0.47, whereas the largest negative

partial correlation is −0.58. The weighted average (using sample size as

weights) of only the negative partial correlations is −0.06, whereas the

weighted average of only the positive partial correlations is +0.07. There is,

however, no reason to separate the studies like this.

5 For studies using the Brown-Medoff methodology, with unionization measured as a percentage, the

total productivity effect is the coefficient on unionism after it is converted into a percentage. In the case

where a union dummy variable is used, the total union productivity effect is calculated as the antilog of

the dummy coefficient with 1 subtracted from it.

Page 11

660 /

C D P L

TABLE 1

E S E A B U P (n = 73)

AuthorCountry

N

Average

t-statistic Average r

δ

Mean

Union

Effect

Total

Union

Effect Outlet

Allen (1984)

Allen (1985)

Allen (1986a)

Allen (1986b)

Allen (1988a)

Allen (1988b)

Argys & Rees (1995)

Arthur (1994)

Baldwin (1992)

Bartel (1994)

Batt (1999)

Bemmels (1987)

Black & Lynch (2001)

Boal (1990)

Bronars et al. (1994)

Brown & Medoff (1978)

Brunello (1992)

Byrne et al. (1996)

Byrnes et al. (1988)

Cavalluzzo & Baldwin (1993)

Caves (1980)

Caves &Barton (1990)

Chezum & Garen (1998)

Clark (1980a)

Clark (1980b)

Clark (1984)

Conte & Svejnar (1988, 1990)

Conyon & Freeman (2002)

USA

USA

USA

USA

USA

USA

USA

USA

Canada

USA

USA

USA

USA

USA

USA

USA

Japan

USA

USA

USA

UK/USA

USA

USA

USA

USA

USA

USA

UK

81

+2.12

+2.31

+1.38

+1.61

+2.79

+2.69

+1.60

+2.24

−1.70

+1.95

−0.52

−1.19

−1.91

+0.54

+1.04

+1.95

−3.23

−0.86

+2.48

+2.13

−1.77

−3.32

−1.75

+2.00

+0.05

−2.33

+2.04

−0.83

+0.244**

+0.237**

+0.253#

+0.190*

+0.223***

+0.470**

+0.029#

+0.416**

−0.141**

+0.160**

−0.038#

−0.108#

−0.078**

+0.035#

+0.039#

+0.139**

−0.103***

−0.075#

+0.241***

+0.239**

−0.270*

−0.202***

−0.019*

+0.195**

+0.025#

−0.034**

+0.170**

−0.068#

+0.15

+0.12

—

—

+0.20

—

—

+0.15

−0.001

+0.42

—

−0.70

—

—

+0.03

+0.16

—

—

—

—

na

−0.005

—

—

—

−0.03

+0.37

—

+5%

+4%

—

—

+6%

—

—

+7%

na

+7%

—

−18%

—

—

—

+5%

—

—

—

—

na

na

—

—

—

−1%

+25%

—

+15%

+12%

+27%

+35%

+20%

+51%

+1%

+15%

−0.1%

+42%

−3%

−70%

−12%

+3%

+3%

+16%

−15%

−11%

+69%

+38%

na

−0.5%

−3%

+10%

−1%

−3%

+37%

−4%

QJE

RES

JLR

ILRR

ILRR

IR

RLE

AMJ

Book

IR

ILRR

ILRR

RES

IRRR

IR

JPE

ILRR

IR

MS

Book

Book

Book

AE

ILRR

QJE

AER

IJIO

Book

102

44

151

306

42

3169

30

167

155

202

46

627

249

670

204

979

128

197

83

47

268

8152

104

465

4681

155

932

Page 12

What Do Unions Do to Productivity? A Meta-Analysis

/ 661

Cooke (1994)

Coutrot (1996)

Craig & Pencavel (1995)

Davies & Caves (1987)

Dickerson et al. (1997)

Eberts (1984)

Edwards & Field-Hendrey (1996)

Ehrenberg et al. (1983)

Fitzroy & Kraft (1987a)

Freeman (1988)

Graddy & Hall (1985)

Grimes & Register (1991)

Guthrie (2001)

Hirsch (1991)

Holzer (1990)

Huselid (1995)

Ichniowski et al. (1997)

Katz et al. (1987)

Kaufman & Kaufman (1987)

Kleiner & Petree (1988)

Kleiner & Ay (1996)

Kleiner & Lee (1997)

Koch & McGrath (1996)

Kurth (1987)

Lee & Rhee (1996)

Lovell et al. (1988)

Machin (1991)

USA

France

USA

UK/USA

UK

USA

USA

USA

Germany

USA

USA

USA

New Zealand

USA

USA

USA

USA

USA

USA

USA

Sweden

Korea

USA

USA

Korea

USA

UK

841

4289

170

+2.59

+2.77

+1.91

−2.05

+0.88

+0.56

+1.36

+0.36

+2.85

+0.86

−1.47

+2.26

+1.00

−6.10

+1.10

+1.00

+1.50

+1.95

−0.64

+2.64

−0.85

−0.13

+0.68

−3.19

−2.33

−2.49

−0.89

+0.090***

+0.048***

+0.152***

−0.236**

+0.086#

+0.010#

+0.146#

+0.024#

+0.260***

+0.034#

−0.193#

+0.050**

+0.090#

−0.077***

+0.196#

+0.180#

+0.032#

+0.380*

−0.114#

+0.120***

−0.183#

−0.010#

+0.039#

−0.464***

−0.196**

−0.486**

−0.063#

—

—

—

—

—

—

—

+29%

+7%

+29%

−13%

+2%

+6%

+18%

9%

+9%

+12%

−11%

3%

13%

−8%

+0.03

+0.1%

+1%

na

−10%

+60%

−40%

−1%

+34%

−8%

−1%

−68%

−13%

ILRR

TE

BP

Book

IRAE

ILRR

RLE

ILRR

QJE

EER

JLR

IR

AMJ

Book

IR

AMJ

AER

BP

JLR

Book

AILR

IR

SMJ

JLR

JLR

JLR

ECO

86

98

−0.133

+0.02

—

—

—

+0.09

+0.12

—

—

+0.13

−0.08

—

+0.001

—

—

—

+0.6

−0.40

—

+0.34

−0.08

−0.01

−0.68

—

+0.2%

—

—

—

+3%

+4%

—

—

+4%

−3%

—

+0.01%

—

—

—

+7%

−34%

—

+7%

−1%

−0.2%

−16%

—

3251

96

256

123

650

60

2062

136

6248

1320

855

2190

33

37

490

29

184

318

50

144

26

208

AuthorCountry

N

Average

t-statisticAverage r

δ

Mean

Union

Effect

Total

Union

EffectOutlet

Page 13

662 /

C D P L

Maki (1983)

Meador & Walters (1994)

Mefford (1986)

Milkman (1997)

Mitchell et al. (1990)

Mitchell & Stone (1992)

Morishima (1990)

Muramatsu (1984)

Noam (1983)

Pencavel (1977)

Register (1988)

Register & Grimes (1991)

Schuster (1983)

Tachibanaki & Noda (2000)

Torii (1992); Torii & Caves (1992)

Warren (1985)

Wilson (1995)

Wilson & Cable (1991)

Canada

USA

USA

USA

USA

USA

Japan

Japan

USA

UK

USA

USA

USA

Japan

Japan

USA

USA

UK

183

889

126

2684

886

+2.47

−1.99

+4.19

+1.38

+1.70

−3.00

+1.00

+1.49

+0.33

−4.09

+3.86

+2.01

+2.35

−3.02

−0.135

−3.12

+0.96

−2.28

+0.182**

−0.067**

+0.360***

+0.029#

+0.084*

−0.331***

+0.131#

+0.094#

+0.010#

−0.501***

+0.250***

+0.058**

+0.259***

−0.091***

−0.013#

−0.583***

+0.112#

−0.146**

+0.33

—

—

—

—

—

+0.001

+0.12

+0.01

−0.22

+0.17

—

—

—

0

−0.81

+0.14

—

na

—

—

—

—

—

33%

−13%

+33%

+19%

+5%

−13%

+0.1%

+12%

+3%

−22%

+19%

5%

17%

−50%

0

−81%

+14%

−18%

RI

JLR

ILRR

JLR

Book

ILRR

IR

Book

RLE

BJIR

JLR

JLR

ILRR

Book

Book

JLR

Book

AE

83

69

+0.05%

+2%

—

−3%

—

—

—

—

0

−19%

+5%

—

515

1100

56

389

1229

474

2358

124

26

266

260

*, **, *** correlation is statistically significant at the 10, 5, and 1 percent levels, respectively. # Not statistically significant. na denotes that the productivity effect cannot be derived

from the study. Journal codes are as follows: AE: Applied Economics; AER: American Economic Review; AMJ: Academy of Management Journal; BJIR: British Journal of

Industrial Relations; BP: Brookings Papers; ECO: Economica; EER: European Economic Review; IJIO: International Journal of Industrial Organization; ILRR: Industrial & Labor

Relations Review; IR: Industrial Relations; JLR: Journal of Labor Research; JPE: Journal of Political Economy; MS: Management Science; RI: Relations industrielles; RES:

Review of Economics and Statistics; RLE: Research in Labor Economics; QJE: Quarterly Journal of Economics; TE: Travail et Emploi.

AuthorCountry

N

Average

t-statisticAverage r

δ

Mean

Union

Effect

Total

Union

EffectOutlet

TABLE 1 (cont.)

Page 14

What Do Unions Do to Productivity? A Meta-Analysis

/ 663

In order to conserve space, only the key and more interesting meta-analysis

results are presented and discussed. The traditional meta-analysis results

are presented in Table 2 for five different groupings of studies. The table

presents information on the number of studies included in each meta-

analysis, the combined sample size of the included studies, and the unweighted

mean, median, and sample size weighted mean partial correlations. The

range shows the spread of actual results reported in the literature. The

95 percent confidence intervals are presented in brackets, and these incorporate

the variance associated with the estimated average partial correlations.

These can be used to test the hypothesis that the union-productivity effect

is zero, positive, or negative. An important consideration is whether the

partial correlations are drawn from a group of studies that is homogeneous.

A chi-square test for this is presented in Table 2, testing the hypothesis that

all the partial correlations are equal. If this hypothesis is rejected, then it is

important to search for factors that lead to heterogeneity.6

There is also the issue of possible differences in the quality between stud-

ies. Our starting position was to rank all the studies equally. The issue of

quality may be reflected in the publication outlet. Of the 73 studies listed in

Table 1, 13 were published in the Industrial and Labor Relations Review, 11

in the Journal of Labor Research, 9 in Industrial Relations, and 13 in leading

economics journals (such as the Journal of Political Economy, American

Economic Review, and the Quarterly Journal of Economics).7 We infer from

this that the published empirical literature is of a very high quality and that

there is no basis for distinguishing articles on the basis of publication out-

let. That is, it is not valid to argue that there is a significant portion of the

studies that were not good enough to get into “good journals.”

As is common in meta-analysis, we use the sample size to construct

weighted means. Thus a study with a larger sample size is given greater

weight regardless of employment levels. There is no way of getting around

this problem because few studies report employment levels. It is not possible

to use employment levels as weights. This problem affects the meta-analysis

but not the meta-regression analysis.8 In addition to using sample sizes as

weights, we also used two alternative weighing methods. The first involved

using citations as weights. Citations were derived from the Social Science

Citations Index. In effect, this is equivalent to using what the profession

6 A technical appendix is available from the authors detailing the formulas (weighted mean, confi-

dence intervals, and the heterogeneity test) used in the meta-analysis. All the meta-analysis calculations

were made using MetaWin version 2. (Rosenberg et al. 2000)

7 This does not imply that the studies published in the other journals are of inferior quality.

8 For comparison purposes, we report both the unweighted and the weighted measures of central

tendency.

Page 15

664 /

C D P L

TABLE 2

M-A U P, P C P E

All Studies (2)U.S. Studies (3)U.K. Studies (4) Japanese Studies (5)

U.S.

Manufacturing (6)

Number of studies

Total sample size

Mean r

73 55 75 10

58 403

+0.03

47 549

+0.05

1 687

−0.17

4 045

−0.01

5 004

+0.12

(−0.21 to +0.26)

+0.03

+0.01

(0.00 to +0.02)

+0.04

(+0.01 to +0.06)

−0.58 to +0.47

511***

(−0.23 to +0.32)

+0.04

+0.02

(+0.01 to +0.03)

+0.06

(+0.03 to +0.09)

−0.58 to +0.47

395***

(−1.00 to +0.75)

−0.15

−0.10

(−0.16 to −0.04)

−0.15

(−0.28 to −0.01)

−0.46 to +0.09

19**

(−1.00 to +1.00)

−0.01

−0.08

(−0.13 to −0.04)

−0.03

(−0.18 to +0.11)

−0.18 to +0.13

28***

(−0.59 to +0.84)

+0.11

+0.07

(+0.04 to +0.10)

+0.10

(+0.01 to +0.20)

−0.20 to +0.42

62***

Median r

Weighted mean r

Random effects weighted mean

Range

Heterogeneity

Productivity Effects

−0.09 [−0.18]Unionization elasiticity

+0.01 [+0.07]

+0.01 [+0.08]na

+0.08 [+0.01]

Total productivity effect

Unweighted

Sample size weighted

Ranking size weighted

+4%

+1%

+7%

+7%

+3%

+7%

−11%

−8%

−14%

−13%

−32%

−13%

+10%

+10%

+2%

N: Figures in parentheses are 95 percent confidence intervals. **, *** Coefficient is statistically significant at the 5 and 1 percent levels, respectively, Chi-square test. na means

sample size too small to calculate average elasticity.