Why do Some Studies Show that GenerousUnemployment Bene�ts Increase UnemploymentRates? A Meta-Analysis of Cross-Country Studies

Jaewon Kim�

This version: January, 2011

Abstract

This paper investigates the hypothesis that generous unemployment bene�tsgive rise to high levels of unemployment by systematically reviewing 34 cross-country studies. In contrast to conventional literature surveys, I perform ameta-analysis which applies regression techniques to a set of results taken fromthe existing literature. The main �nding is that the choice of the primary dataand estimation method matter for the �nal outcome. The control variables inthe primary studies also a¤ect the results.JEL-codes: C42, J65.Keywords: meta-analysis, cross-country study, unemployment, bene�t re-

placement rate, bene�t duration.

1 Introduction

There has been a popular belief held by many economists and in�uential organisa-

tions that generous unemployment bene�ts and other labour market institutions are

responsible for high and persistent unemployment. In particular, the OECD argued

�Department Economics, Stockholm University 106 91 Stockholm Sweden. I thank Ann-So�eKolm, Matthew Lindquist, and Helena Svaleryd for their advice and support. I also thank seminarparticipants at Stockholm University, Uppsala University, and SOFI for valuable comments andsuggestions. E-mail: Jaewon.Kim@ne.su.se

1

in the OECD Employment Outlook (1996, 1997, 1999) and the OECD Jobs Strategy

(1996) that so-called rigidities imposed by labour market institutions increased un-

employment. It suggested labour market deregulation as a remedy for unemployment

problems in the member countries. The IMF also argued that generous unemploy-

ment insurance systems have contributed to high unemployment (World Economic

Outlook, 2003). More recently, Jean-Claude Trichet, president of the European Cen-

tral Bank, acknowledged the need of labour market �exibility in his speech.1

Over the last two decades, economists have analysed the relationship between

labour market institutions and unemployment using cross-country regression meth-

ods. It appears, however, that no clear consensus over the empirical evidence has

been reached. For example, Nickell et al. (2001) and Nunziata (2002) found a highly

signi�cant positive correlation between the unemployment bene�ts and the unem-

ployment rate, while Baker et al. (2003) found no clear relationship. Moreover, Belot

and van Ours (2001, 2004) reported that the generosity of the unemployment bene-

�ts is negatively correlated with unemployment in several speci�cations. What is it

that leads these economists to such di¤erent results? Can we explain variation in the

results by their di¤erent choices of data, estimation methods, and model speci�ca-

tions? Will these studies reach an unanimous conclusion once one takes account of

the di¤erences in the empirical settings?

To answer these questions, this paper presents a meta-analysis of existing cross-

country studies which look at the impact of labour market institutions on the aggre-

gate unemployment rate. More speci�cally, it focuses on the question of whether or

not generous unemployment bene�ts, in terms of high replacement rate and lengthy

durations of bene�ts, create high unemployment. There are a few forerunners who

evaluate the robustness of the results of the e¤ects of labour market institutions.1The speech by Jean-Claude Trichet, Amsterdam, 15 February 2007.

2

Baker et al. (2003) compared seven papers and acknowledged the lack of robustness

in the estimates. Bassanini and Duval (2006) showed how the estimates can vary

depending on empirical speci�cations and the choice of the data.

This paper is the �rst to assess the link between generosity of unemployment

insurance systems and the level of unemployment using a meta�analysis. A meta-

analysis enables us to probe whether the variation in empirical outcomes found across

the studies relies on the use of di¤erent regression methods, how the data are de�ned

or organised, or which control variables are used in the primary studies. Identifying

the e¤ects of di¤erent empirical settings is essential for applied economists and policy

makers when discussing the disputed issue of the e¤ect of unemployment bene�ts on

the aggregate unemployment rate.

The meta-analysis is based on a specially constructed meta-data set by reviewing

34 studies that empirically analyse the relationship between measures of unemploy-

ment bene�ts and the level of unemployment. This paper �nds that the choice of the

primary data, data de�nition of unemployment bene�ts variables, and empirical spec-

i�cation have a considerable in�uence on the �nal outcomes. The included control

variables in the primary studies also a¤ect the results whether the studies reached a

positive or no relationship between unemployment bene�ts and unemployment rates.

The remainder of this paper is organised as follows. In the following section, I dis-

cuss the existing theoretical and microdata-based empirical literature concerning the

relationship between unemployment bene�ts and unemployment. Section 3 describes

the methodology of meta-regression analysis. In section 4, I present the meta-data set.

The results from the baseline estimations and their sensitivity analysis are discussed

in section 5. Section 6 concludes.

3

2 Theories and Empirical Studies based on Micro-

data

Before beginning the meta-analysis of the cross-country studies which use aggregate

data, I take a brief look at the existing theoretical as well as micro-empirical literature

on the relationship between the generosity of unemployment insurance systems and

unemployment rate.

A highly in�uential theoretical work on how the generous unemployment insur-

ance systems a¤ect individual unemployment spells was done by Mortensen (1977).

He argues that since unemployment compensation reduces the income foregone while

currently unemployed but does not a¤ect future income, the expected search duration

increases with the bene�t rate. At the same time, a worker either currently employed

or unemployed and not receiving bene�ts will be eligible to receive the bene�ts during

any future employer initiated unemployment spell. Consequently, an improvement

in either the bene�t rate or the maximum bene�t period makes current employment

relatively more attractive. In response, an individual who is unquali�ed to receive

the bene�t such as new entrants, exhaustees and those who quit will �nd employ-

ment more quickly by both lowering his/her reservation wage and by searching more

intensively. There is, thus, no strong theoretical reason to expect an unambiguous

relationship between the generosity of unemployment bene�ts and individual unem-

ployment spells.

Thanks to the development of rich micro-data, a large number of empirical studies

have been done. The results in these studies are mixed. To name a few, Jones

(1995) found from the Canadian unemployment insurance system that the cut in

wage replacement rates is associated with longer individual unemployment spells.

Carling et al. (2001) found that a decrease in the replacement rate in the Swedish

4

unemployment insurance from 80% to 75% caused an increase in the transition rate

from unemployment of roughly 10%. Similarly, Roed and Zhang (2003) show using

Norwegian data that a marginal increase in compensation reduces the escape rate

from unemployment signi�cantly.

Despite the large number, these empirical studies based on micro-data have some

weakness. The fundamental limitation is due to their partial equilibrium nature. As

Holmlund (1998) acknowledged, the micro-data analyses on unemployment duration

are of only limited use, as they do not capture the general-equilibrium e¤ects. From

a policy perspective, it is valuable to identify how di¤erent unemployment bene�t

regimes may result in the di¤erence in the level of unemployment over a longer time

period. Theoretically, for example, we know that more more generous unemployment

bene�ts may a¤ect wage formation and induce higher wage demands, which can

increase the long-run unemployment rate. Here is where the cross-country studies

reviewed in this paper have an important role to play. However, as we shall see, these

studies have not yet reached a consensus.

3 Methodology

In this section, I describe the meta-analysis method. In contrast to conventional

literature surveys, a meta-analysis applies regression techniques to a set of results

taken from the existing literature to summarise the results on particular topics, to

provide an aggregate overview of a subject, and to allow an analysis of factors that

may in�uence the results. The main idea of this meta-analysis is that the coe¢ cient

estimates of unemployment bene�ts in unemployment regressions are constant across

the studies if one successfully controls for the di¤erent settings found in the empirical

studies. It enables us to quantify the e¤ects of particular choices of data, estimation

5

methods, or model speci�cations for the outcomes in the primary studies.

The procedure of this meta-analysis is as follows. First, I search for empiri-

cal papers that investigate the e¤ect of unemployment bene�ts on unemployment

rate with cross-country data. Second, I review the selected papers and construct a

meta-data set that consists of meta-dependent and meta-independent variables. The

meta-dependent variables are the coe¢ cient estimates of unemployment bene�ts in

unemployment equations drawn from regressions reported in the primary studies.

The meta-independent variables are the binary variables that characterise the em-

pirical settings of each regression. Third, I conduct meta-regression analysis using

probit models and OLS.

The meta-observations are obtained from existing journal articles and other pub-

lished or unpublished papers. I search for published articles using the search engine,

googlescholar, with search words "labour market institutions" & "unemployment"

or "unemployment bene�t j unemployment insurance" & "unemployment". Unpub-

lished working papers, dissertations, essays, and mimeographs are found by the search

engine google.2 References in each paper are cross searched until no new studies are

found. I consider only the papers that are from the year 1988 and up until march,

2007, when this study was started. This restriction, however, does not exclude any

earlier relevant studies because most of the studies are done after the year 1996, when

the extensive panel data set of labour market institutions became available.

To be an observation point in this meta-analysis, a study is required to have

(1) an empirical analysis of the cross-country data of the OECD countries, (2) well-

de�ned and well-reported source of the data, and (3) the estimates derived from a

regression analysis with a dependent variable measuring the level of the aggregate

2Although previous meta-studies in other topics used other computerised data bases as IDEASand RePEc, EconLit, in recent years, googlescholar comprises all above databases.

6

unemployment rate and independent variables any among three measures of generos-

ity of unemployment bene�ts, i.e. unemployment bene�t replacement ratio, bene�t

duration, or an interaction of these two. I also include regressions when their de-

pendent variables are long-term or youth unemployment rate.3 A large fraction of

articles were either theoretical or micro-data analyses and discarded from this analy-

sis. Studies that analysed only one or a few countries, or each country separately are

excluded. Finally, 34 papers that ful�ll the requirement are included in this meta-

analysis. When more than one regression is reported, I use all estimates that meet

the above requirements. Taking multiple estimates from an article is preferable to

collecting a single value only from each study, because procedures representing each

study by a single value result in a serious bias and a loss of information (Bijmolt and

Pieters, 2001).4

The meta-data are �rst estimated using a probit model. Employing a probit

model for a meta-analysis in this fashion is novel. The goal of the �rst experiment

is to identify the factors that produce signi�cant positive correlations between the

generosity of unemployment bene�ts and the level of unemployment. For this exper-

iment, I construct the binary meta-dependent variables.5 The probit meta-regression

is

Pr(bj = 1jZj1; Zj2; :::; ZjK) = G(� +KXk=1

�kZjk); j = 1; 2; :::L,

where bj is the binary meta-dependent variable constructed from the reported es-

3The studies that analysed male and female unemployment rate separately such as Jimeno &Rodriguez-Palenzuela (2002) are not included in this meta-analysis.

4The e¤ect of publication bias is partially mitigated by sampling all estimates in each primarystudy and by including both published and unpublished studies.

5If the reported estimate of the measures of unemployment bene�ts from the primary equationsis signi�cantly positive at the 10% level, the binary meta-dependent variable gets the value 1.Otherwise, it gets the value 0. I also conducted the probit meta-analysis with the binary meta-dependent variables with the signi�cant level of 5% and 1%. The estimated results are consistentwith those of the 10% level.

7

timates in the jth regression in the literature comprised of L regressions, � is the

summary value of the marginal e¤ect b, the Zjk�s are the meta-independent variables

which measure relevant characteristics of the observation, and the �k�s are the meta-

regression coe¢ cients which re�ect the e¤ect of particular regression characteristics.

G is the standard normal c.d.f. Since the observations within paper are more likely

to share similar characteristics, the data are estimated with robust standard errors

for clustered samples by paper.

Following this, I run an experiment in which I use the coe¢ cient estimates them-

selves as the dependent variables. The goal of this second experiment is to quantify

factors that give rise to large correlations between the generosity of unemployment

bene�ts and the level of aggregate unemployment. The meta-regression model is

estimated using OLS,

bj = � +KXk=1

�kZjk + ej; j = 1; 2; :::L,

where bj is the reported estimate and � the summary value of the e¤ect of generosity

of unemployment bene�ts in unemployment rate. The meta-regression coe¢ cient �k

re�ects the marginal e¤ect of particular regression characteristics. The error term

ej is robust for clustered samples by paper.6 In the sensitivity analysis, I include

the estimation by weighted least squares (WLS), where each coe¢ cient estimate is

weighted by the inverse of its standard error.7

6Jeppesen et al. (2002) and Disdier & Head (2008) suggest random e¤ects panel speci�cation todeal with the multiple estimates from each paper. The random e¤ects model places greater emphasison within-paper variation than cross paper variation. The Breusch-Pagan tests imply that theresearcher random e¤ects are insigni�cant for the coe¢ cient estimates of duration of unemploymentbene�ts (BD) and the interaction (BRRBD). The existence of the random e¤ects is ambiguous forthose of bene�t replacement rate (BRR). This suggests that researchers in this literature are notconducting research in a manner fundamentally di¤erent from one another. Hence, I use the OLSand probit model for the baseline regressions.

7The WLS is one of the common practice in meta-regression analysis, when one is interestedin the particular e¤ect size. Since the purpose of this paper is to assess the factors that give rise

8

4 Data

This section presents the meta-data in detail. In primary studies, the generosity

of unemployment bene�ts is often de�ned as the level of bene�t replacement rate,

its duration, or the interaction of replacement rate and duration of unemployment

bene�ts. Hence, I conduct the meta-analysis on these three dependent variables;

the coe¢ cient estimates of bene�t replacement rate (BRR), those of duration of

unemployment bene�ts (BD), and those of the interaction of these two (BRRBD) in

the unemployment equations. For probit estimation, I construct the binary variables

from these coe¢ cient estimates. The binary meta-dependent variable BRR10 is equal

to 1 if the coe¢ cient estimate of the bene�t replacement rate in a primary equation

is signi�cantly positive at the 10% level. The 34 papers give 382 observations for

bene�t replacement rate, 111 for bene�t duration, and 40 for the interaction of these

two. About half of the observations of the bene�t replacement rate and its duration

are signi�cantly positive at the 10% level. About 80% of the 40 observations of

the interaction term BRRBD are signi�cantly positive. A substantial number of the

observations is from the primary equations where the left-hand side variable is the

natural logarithm of unemployment rate. For more comparable results, I antilog the

values from these primary equations by multiplying the natural logarithm base e.

The descriptive statistics of the respective meta-dependent variables are presented in

Table 1 in Appendix A.

Figures 1 and 2 illustrate how the reported estimates of the bene�t replacement

rate and the duration of unemployment bene�ts are scattered over years of publi-

cation.8 The dispersion in the literature on the relationship between the generosity

to signi�cant positive coe¢ cient estimates or larger estimates rather than to �nd the e¤ect size, Iinclude the WLS estimation as robustness check.

8For working papers and other unpublished papers, I use the year when the latest version iswritten.

9

of unemployment bene�ts and the level of unemployment seem to have increased

over time. Although the positive estimates dominate, there appear more negative

estimates in the recent papers.

The meta-independent variables Zjk are dummy variables which characterise each

observation; see Table 2. I divide these variables into four groups; variables for (A)

data, (B) estimation methods, (C) model speci�cations of the primary studies, and

(D) others. The �rst category of the meta-independent variables de�nes di¤erences in

the data used in the primary studies. Across the meta samples, two measures of ben-

e�ts replacement rate are used. The �rst measure de�nes bene�t replacement rate as

the �rst year of unemployment bene�ts, averaged over three family situations (single,

with dependent spouse, with spouse in work) and two earnings level (100% and 67%

of APW earnings). The data are provided by Nickell, who modi�ed and completed

the OECD data set. A large fraction of the primary studies used the data with this

de�nition. The other is the OECD summary measure of bene�t entitlements, which

also takes unemployment durations into account.9 The latter is used in Elmeskov et

al. (1998), Belot and van Ours (2001, 2004), Boone and van Ours (2004), Macculloch

and di Tella (2002), and Kenworthy (2002) and partly in Scarpetta (1996) and Bas-

sanini and Duval (2006). The meta-independent variable ubsum=1 is for the paper

that used the OECD summary data of bene�t entitlements, and ubsum=0 for the

paper that used the bene�t replacement rate data by Nickell.

The second meta-independent variable for data is bdyear, which is used for the

meta-dependent variables, bene�t duration (BD) and the interaction between re-

placement rate and duration (BRRBD). Throughout the literature, unemployment

bene�t duration is quanti�ed either as the maximum duration of the unemployment

9It is de�ned as the average unemployment bene�t replacement rate across two income situations,three family situations, and three di¤erent unemployment durations (1st year, 2nd and 3rd years,and 4th and 5th years of unemployment).

10

insurance bene�ts in months or years, or as an index constructed by Nickell.10 The

studies done after the year 2002 often use the index, while earlier studies used bene�t

duration expressed in years. The variable avg5 is equal to 1 if �ve-year averaged data

instead of yearly data are used.11 About 1/3 of the primary studies used �ve-year

averaged data.

The variables, lterm and youthun, indicate whether the primary regressions analyse

long-term respective youth unemployment rates. Besides total unemployment rate,

Nickell (1997 & 1998) analyses long- and short-term unemployment rates. Scarpetta

(1996) analyses long-term and youth unemployment rates. Esping-Andersen and

Regini (2000) also use youth unemployment rate as their dependent variable.

The meta-independent variable for data, timinvinst, is for if the studies used time-

invariant data of labour market institutions. Blanchard and Wolfers (2000), Chen et

al. (2003), Algan et al. (2002), as well as, earlier studies such as Jackman et al.

(1990) and Burda (1988) used time-invariant labour market institutions data.

The number of the countries that are included in the primary studies is limited to

eighteen to twenty throughout the literature due to the availability of the data.12 The

exceptions are Macculloch and Di Tella (2005) which has 21 countries by including

Greece, and Jackman et al. (1990) that have only fourteen countries.13 The country

dummy variables portugal and spain in the baseline equations and austria, ireland,

newzealand, spain, and switzerland in the robustness tests control for inclusiveness of

the respective countries. These countries have been occasionally disregarded in the

10The bene�t duration index by Nickell is de�ned as bd = 0:6� brr23brr1 +0:4�brr45brr1 , where brr1, brr23;

and brr45 refer to the �rst, second- and third-, and 4th- and 5th-year of gross bene�t replacementrates.11I code 4 observations that use 6-year averages to avg5=1.12These 20 countries are Australia, Austria, Belgium, Canada, Denmark, Finland, France, Ger-

many, Ireland, Italy, Japan, Netherlands, New Zealand, Norway, Portugal, Spain, Sweden Switzer-land, United Kingdom and United States.13The studies that had a smaller number of countries such as Stockhammer (2004) which had the

data of 5 countries are discarded from the meta sample.

11

analyses due to the unavailability of earlier data. About one third of the primary

studies do not include Portugal.

The variable year6090 de�nes if the primary study uses the data of the time

period between the year 1960 and the 1990s. Many studies written after the year

1999 used data from 1960 to the 1990s, because the CEP-OECD Institutions Data

Set became available. Earlier studies such as Nickell (1997 & 1998), Scarpetta (1996),

and the OECD (1999) take the time period of the 1980s and 1990s. The variables,

time60, time70, time80, and time90 indicate if the respective decades are included

in the primary analyses. About 40% of the selected primary studies used the time

periods from 1960 to sometime in the 1990s.

The second category of the meta-independent variables is estimation method.

Over 90% of the observations are estimated by the linear regression models such as

OLS, random-e¤ects, and �xed-e¤ects model. Non-linear least squares are used only

in a few studies, e.g. in Blanchard and Wolfers (2000), partly in Bertola et al. (2001),

Bassanini and Duval (2006), and Macculloch and Di Tella (2005). The ols, re, fe,

and nonlinear control for the use of the respective regression methods in the primary

studies.

The third group is the variables that describe model speci�cation. All primary

equations are unique in their choice of the explanatory and the control variables. In

general, labour market institutions such as union density, degree of coordination be-

tween union and employer in wage bargaining, employment protection legislation, as

well as tax on labour and tax wedges are often included. The meta-independent vari-

ables, ud , coord, ep, tax, and twedge describe the inclusiveness of these explanatory

variables.

The control variables vary from study to study. The macroeconomic variables

that can be argue to a¤ect aggregate unemployment are often included in the primary

12

studies. For instance, an adverse labour shock such as a decline in the gap between

the wage rate and the marginal product of labour, or a shift in production techniques

away from labour and towards capital may raise unemployment. A rise in real interest

rates a¤ects negatively capital accumulation and productivity, thereby reduces labour

demand at a given wage level and increases unemployment (Blanchard & Wolfers,

2000). Thus, I include the use of the macroeconomic controls such as the changes

in in�ation, labour demand shocks, total factor productivity shock, home occupation

rate, and real interest rate in the primary literature. The meta-explanatory variables

in�a, lds, tfps, home, and rinterest indicate whether these variables are controlled.

Moreover, �rstdi¤=1 signi�es the use of the �rst-di¤erence of unemployment rate.

Baccaro and Rei (2005), Fitoussi (2000), Daveri and Tabellini (2000), and Macculloch

and Di Tella (2005) had several speci�cations of �rst-di¤erence model. There is no

observation of bene�t duration that has the �rst-di¤erence model. Also, lagun indi-

cates the use of an auto-regressive model. The studies that use annual observations

often allow for an auto-correlation of one-period.

Finally, I control for the quality of the results by adding a dummy "rank" which

equals to 1 if the study was published in the journals with the rank 3 or above in

"Keele Four-Four-Two list: Ranking of Economics Journals" and 0 otherwise. Among

the quality journals with the rank 4 the primary studies are published in European

Economic Review or Journal of Economics Perspectives. Some studies are published

in the rank 3 journals such as Economic Journal, Oxford Economics Paper, Brookings

Paper of Economic Activity, and Economic Policy. The idea is that the standard for

methodological rigor could be higher at quality journals and this could correct biases

present in other estimates (Disdier and Head, 2008).

13

5 Meta-Regression Analysis

This section presents the results of the meta-regression analysis. Section 5.1 presents

the baseline estimation results. Section 5.2 reports the results of the sensitivity

analysis. In section 5.3, I discuss the �ndings from this meta-regression analysis.14

5.1 The Baseline Results

Table 3 presents the baseline probit estimation for the binary variables of bene�t

replacement rate (BRR10) and bene�t duration (BD10).15 The choice of data a¤ects

whether or not the primary studies obtain signi�cantly positive estimates on both

bene�t replacement rate and bene�t duration. The studies that used the OECD sum-

mary measure of bene�t entitlements (ubsum) and the bene�t duration measured in

years (bdyear) have a higher probability of getting signi�cantly positive estimates.

Studies that use �ve-year averaged data (avg5 ) are less likely to have positive esti-

mates on unemployment bene�t replacement rate, but more likely to have positive

estimates on bene�t duration.

The estimation method and the choice of the control variables also matter. Fixed

e¤ects model (fe) tends to give more signi�cantly positive estimates of the bene�t

duration. The primary equations that control for labour demand shocks (lds) are

more likely to obtain positive correlations for the bene�t replacement rate, but this

probability decreases for the bene�t duration estimates. The owner occupation rate

14The issue of publication bias, which is the tendency among editors of academic journals topublish results that are statistically signi�cant, is beyond the scope of the current study.15The results of the probit estimation with BRR5, BRR1, BD5, and BD1, which are the binary

meta-dependent variables indicating 1 for signi�cant positive estimate in 5% level, respective 1%and 0 for otherwise, are similar to those in Table 3. Due to the small number of observations, theprobit estimation of the interaction meta-dependent variable (BRRBD10) is excluded. I also appliedlinear probability model on these binary data. The results are consistent with those by the probitmodel. The linear probability model neither gives any meaningful results for BRRBD10 due to thesmall number of observations.

14

(home), which Oswald (1999) argued as an important control in the unemployment

rate-labour market institutions equations, works in opposite direction that it gives

lower probability of the positive and signi�cant bene�t replacement rate estimates,

while the bene�t duration estimates are more likely to be signi�cant and positive in

the primary unemployment equations.

Table 4 shows the OLS estimation for the coe¢ cient estimates of bene�t replace-

ment rate (BRR), bene�t duration (BD), and the interaction between these two

(BRRBD). As a baseline approach, the size of the reported coe¢ cient estimates are

analysed regardless of their signi�cance or sign. Using the OECD summary data

gives a negative biasing e¤ect to the size of the coe¢ cient estimates of BRR.16 The

studies that use the bene�t duration measured in years tend to relatively overesti-

mate the e¤ect of BD and underestimate that of BRRBD. Moreover, using the data

that are organised in �ve-year averages has a negative bias on the BRR and BRRBD

estimates, while this e¤ect is insigni�cant for BD.

When long-term unemployment rate or youth unemployment rate are analysed in

the primary equations, the e¤ect size of the bene�ts on the level of unemployment

prones to be smaller. This implies that youth population or long-term unemployed

are less likely a¤ected by generous unemployment replacement rate. It re�ects the

fact that in most countries the unemployment insurance systems are often directed to

the incumbents of the labour market who have recently lost job, but not for those who

newly entered the labour market or who are unemployed for longer period. Similar

to the baseline probit estimation, the macroeconomic controls or other institutional

variables have considerable e¤ects. For example, labour demand shocks give positive

16Unlike the conventional de�nition of bias in the econometric literature, in this paper "posi-tive/negative bias" implies that the particular meta-independent variable contributes to the resultsabove/below the average, which is predicted probability in the probit estimation or the constant inthe OLS estimation.

15

bias on the coe¢ cient estimates of BRR, but negative bias on the interaction term

(BRRBD).

The highly signi�cant positive intercepts of BRR and BRRBD can be interpreted

as the summary for the primary studies that have all meta-independent variables

equal to zero. However, these constant terms are relatively small, which means

that they are likely to be zero or negative once one starts adding alternative meta-

independent variables. The primary studies that found a positive correlation between

unemployment bene�ts and unemployment rate can, hence, easily be altered by choos-

ing alternative empirical speci�cations.

In all baseline equations of the probit and OLS models the dummy variable "rank"

is insigni�cant except in the OLS model of the interaction BRRBD. It suggests that

the signi�cance or the size of how the overall generosity of unemployment bene�ts

a¤ects the aggregate unemployment rate do not seem to be dependent on whether the

studies are published in the quality journals or not. The high values on the diagnostic

statistics, Wald test �2 and pseudo-R2 in the probit model and F-statistic and R2

in the OLSmodel, indicate that these baseline models are signi�cant and perform well.

5.2 Sensitivity Analysis

To test the robustness of the baseline results, a number of alternative meta-explanatory

variables describing the use of the primary data, estimation methods, and control vari-

ables are introduced. Various speci�cations and subsets of the observations are also

probed. Besides the probit model and the OLS, I run the meta-regressions using the

weighted least squares (WLS) with weights equal to the inverse of the standard er-

rors of each observation. Overall the baseline results are robust under the alternative

16

meta-explanatory variables, model speci�cations, and estimation method.

Table 5 presents the sensitivity analysis for the coe¢ cient estimates of bene�t

replacement rate. In columns (1) and (3), I introduce additional country dummies,

austria, ireland, newzealand, spain, and switzerland. The dummy for time period of

the data, year6090, is divided into ten-year as time60, time70, and time80. In column

(2), I use additional meta-independent variables describing estimation methods, re,

ols, and nonlinear, as well as, control for other labour market institution variables,

employment protection (ep) and union coverage (uc). I also control for the year of

publication of each study, which is non-binary and centred (publication). The highly

signi�cant negative coe¢ cient estimate on the publication year re�ects what we saw

earlier in Figure 1 and 2; namely that this literature tends to deviate more as time

progresses from the early consensus concerning the positive correlation between the

generosity of unemployment bene�ts and unemployment rate. In column (4), only

the observations that are signi�cant at the 5% level are considered. In column (5), the

baseline speci�cation is estimated by the WLS. The sensitivity analysis results of the

BRR estimates are robust, except the constant term became marginally insigni�cant

when the WLS model is used.

Table 6 shows the sensitivity analysis for the bene�t duration (BD) and the inter-

action (BRRBD). In columns (1) and (3), the alternative dummy variables describing

inclusiveness of countries, time period of the data and estimation methods of the pri-

mary regressions are used. In column (2), I control for the use of other labour market

institutions in the primary equations, as well as, publication year. In column (4), only

the reported BD estimates that are signi�cant at the 5% level are considered. The

baseline speci�cation for the BD is estimated by the WLS model in column (5). Col-

umn (6) and (7) are the results of the robustness checks for the BRRBD. In column

(6), the alternative variables for time of the data and publication year are speci�ed.

17

The last column is the meta-regression results for the BRRBD coe¢ cient estimates

using the WLS model.

Throughout the sensitivity tests, the biasing e¤ect of the use of long-term and

youth unemployment rates, �ve-year averaged data, �xed e¤ects model, and the con-

trol variables such as labour demand shocks, total factor productivity shocks, the

degree of coordination, and the owner occupation rate remain to be highly signi�-

cant. The choice of primary variable, bdyear, may become marginally insigni�cant in

a few speci�cations. The negative coe¢ cient estimates on the publication year again

indicate the dissension in the literature over time.

The sensitivity analysis can be summarised as follows. First, the biasing e¤ects of

di¤erent data choice are in general robust except in a few cases of marginal insignif-

icance. Second, the e¤ects of using �xed e¤ects model and some control variables

remain highly robust for the bene�t duration (BD) and the interaction between re-

placement rate and duration (BRRBD). Third, over time the literature of the e¤ects

of generous unemployment insurance systems on unemployment rates seems to di-

verge from the positive correlation to the negative or no signi�cant correlation. As

can be seen in Figure 1 and 2, the literature is actually moving away from a consen-

sus. This is one of the main motivation for the type of meta-analysis presented in

this paper.

5.3 Discussions

The meta-analysis of the e¤ects of unemployment insurance system on unemployment

rates suggests the following �ndings. First, the choice of the primary data a¤ects the

size of the reported estimates and the probability of obtaining signi�cantly positive

estimates. For example, the OECD summary measure of unemployment bene�t en-

18

titlement and the bene�t duration measured in years bias downward, respective, up-

wards. These two increase the probability of obtaining signi�cant positive estimates

in the primary equations. Comparsion across the OLS and the probit estimation

results is, however, not straightforward, since signi�cance is a function of two pa-

rameters, i.e. point estimates and standard errors, whereas magnitude is a fuction

of one. A positive coe¢ cient in Table 3 and a positive coe¢ cient in Table 4 are

roughly consistent in the sense of unambiguously producing evidence in favor of the

unemployment increasing e¤ect of the generosity of unemployment bene�ts.

Can we assess whether or not one of these data sets is better than the other? When

we compare the bene�t replacement rate data by Nickell and the summary bene�t

entitlements data by OECD, the values of the later data are lower than the former

except the case of Australia. The studies that used Nickell�s bene�t replacement rate

alone contain insu¢ cient information on the generosity of unemployment insurance

system, since the OECD summary data takes information about the duration of the

bene�ts into account. When it comes to the data for bene�t duration, the index by

Nickell captures the level of bene�ts available in the later years of a spell relative

to those available in the �rst year, while the bene�t duration expressed in years

only tells the length of the duration. The BD-index by Nickell is, thus, able to

distinguish the situation in which the level of the unemployment bene�ts decreases

over time from that when the bene�ts do not change over time. The duration index

by Nickell contains more information about the generosity of unemployment bene�ts.

Neverthless, these indices are far from perfect in describing reality, as Blanchard

(2007) acknowledged, "The problem is with the crude measures of institutions, not

with in unemployment".

Second, for the bene�t duration and the interaction between bene�t duration and

bene�t replacement rate, �xed e¤ects model increases the probability of obtaining

19

signi�cantly positive estimates as well as the size of the estimates. In other words,

the studies that estimated by �xed e¤ects model prone to show stronger evidence

in support of the positive association between generous unemployment bene�ts and

high unemployment rates. A large number of the selected studies used panel data,

which are often estimated by �xed e¤ects model, random e¤ects model, or pooled

OLS. While the �xed e¤ects model only allow time-variant independent variables,

the pooled OLS can have time-invariant labour market institution variables, which

are used in earlier studies. However, with the pooled OLS, the country-speci�c ef-

fects are restricted to zero, which is an unreasonable assumption for the cross-country

studies. In addition, the �xed e¤ects model is better than the random e¤ects model

for cross-country studies, because the former is able to control for country-speci�c

unobserved heterogeneity. For a random e¤ects model, one needs to make an addi-

tional assumption that the individual country e¤ects are randomly distributed. With

a small number of observations, such as twenty countries, all within the OECD, this

assumption is unlikely to hold. The �xed e¤ects model is, hence, arguably more

suitable for analysing this question.

Finally, the meta-data shows that whether the primary studies are published in the

quality journals or not does not determine if the primary estimates are signi�cantly

positive or have higher coe¢ cient estimates. It instead seems to depend on di¤erent

choices of empirical setting. There could be certain estimation methods and data sets

that are argued to be better than others. The current meta-study, however, does not

assess an analysis on which empirical choice is made in the quality journals.

20

6 Conclusion

The purpose of this study has been to explain why the numerous cross-country studies

on the relationship between the generosity of the unemployment insurance system and

the level of unemployment have reached such di¤erent conclusions. Using a meta-

regression analysis, it is demonstrated how previous studies that use a particular

set of data, estimation methods, or model speci�cations have a higher probability of

obtaining signi�cantly positive results or larger coe¢ cient estimates. Based on the

�ndings from this meta-analysis, can we now say something more about why these

studies have reached such di¤erent results?

Since the empirical results of the primary studies are the product of several po-

tentially biasing factors, it is not always easy to see how the outcome of a study is

directly related to the empirical setting which it adopts. For example, Nickell et al.

(2001 & 2005) and Nunziata (2002) conclude that generous unemployment bene�ts

give rise to higher unemployment. They use �ve-year averaged data of the period

1960-95 for twenty countries and control for changes in in�ation. These choices are

likely to reduce the probability of obtaining a positive relationship between unemploy-

ment bene�ts and unemployment rates. However, they include Portugal and control

for total factor productivity shocks. These two factors can give a higher probabil-

ity of obtaining a positive association between generous unemployment bene�ts and

unemployment rates.

On the other hand, several studies that produce negative or insigni�cant associ-

ations can be more readily explained by the �ndings from this meta-analysis. Baker

et al. (2003) use the twenty-country data set with �ve-year averages. They estimate

the data using random e¤ects models. They also add a control for changes in in-

�ation. According to the meta-analysis, the combination of these choices may have

21

contributed to their negative or insigni�cant estimates. A similar line of reasoning

can be applied to the studies of Belot and Van Ours (2001 & 2004). Excluding Portu-

gal, controlling for in�ation, and a use of �ve-year averaged data can have increased

the probability of obtaining negative and/or insigni�cant results.

This thought-experiment leads us to a corollary. The positive relationship between

generous unemployment bene�ts and the level of unemployment found in the studies

reported here is not readily explained using so-called the biasing factors uncovered

by this meta-analysis, while the negative and/or insigni�cant relationship is more

readily explained by these factors. This may indicate that generous unemployment

bene�ts in terms of high wage replacement ratio may indeed increase the level of

unemployment in cross-country studies considered in this meta-analysis.

Despite increasing dissent in the empirical results in the literature on the relation-

ship between labour market institutions and unemployment, there are still needs of

more studies. The future research may involve not only the idea that unemployment

rate is determined by labour market institutions, but also how the structure of labour

market institutions are a¤ected over time by the level of unemployment, as well as

other economic factors. To do so, improvement of the labour market institutions data

should be preceded.

22

References

[1] Bijmolt, T. and R. Pieters (2001), "Meta-analysis in marketing when studiescontain multiple measurements," Marketing Letters, 12 (2), 157-169.

[2] Blanchard, O. (2007), "A Review of Richard Layard, Stephen Nickell, andRichard Jackman�s Unemployment: Macroeconomic Performance and theLabour Market," Journal of Economic Literature, 45 (2), 410-418.

[3] Carling K., B. Holmlund, and A. Vejsiu (2001), "Do Bene�t Cuts Boost JobFinding? Swedish Evidence from the 1990s," The Economic Journal, 111 (474),766-790.

[4] Disdier, A-C. and K. Head (2008), "The Puzzling Persistence of the DistanceE¤ect on Bilateral Trade," The Review of Economics and Statistics, 90 (1), 37-48.

[5] Holmlund B. (1998), "Unemployment Insurance in Theory and Practice," Scan-dinavia Journal of Economics 100 (1), 113-141.

[6] Jeppesen, I., J. A. List, and H. Folmer (2002), "Environmental Regulations andNew Plant Location Decisions: Evidence from a Meta-Analysis," Journal ofRegional Science, 42 (1), 19-39.

[7] Jones, S. (1995), "E¤ects of Bene�t Rates Reduction and Changes in Entitlement(Bill C-113) on Unemployment, Job Search Behaviour and New Job Quality,"Human Resource Development Canada, August.

[8] Mortensen, D. T. (1977), "Unemployment Insurance and Labor Supply Deci-sions," Discussion Papers 271, Northwestern University, Center for MathematicalStudies in Economics and Management Science.

[9] Roed K. and T. Zhang (2003), "Does Unemployment Compensation A¤ect Un-employment Duration?," The Economic Journal 113 (484), 190-206.

[10] Stanley, T. D. (2005), "Beyond Publication Bias," Journal of Economic Surveys,19 (3), 309-345.

[11] Stanley T. D. and S. B. Jarrell (1989), "Meta-regression Analysis: A Quan-titative Method of Literature Surveys," Journal of Economic Survey, 19 (3),161-171.

[12] World Economic Outlook (April, 2003), Growth and Institutions, World Eco-nomic and Financial Surveys, International Monetary Fund.

23

Appendix A Tables and Figures

Table 1: The description of the meta-dependent variablesBinary Description Mean Std. Min Max # obsBRR10 1=BRR, positive at 10%. 0.534 0.499 0 1 382BD10 1=BD, positive at 10%. 0.523 0.502 0 1 111

Nonbinary Mean Std. Min Max # obsBRR Bene�t replacement rate 0.528 1.359 -2.14 10.72 382BD Bene�t duration 0.225 1.275 -6.685 3.955 111

BRRBD Bene�t replacement rate*duration 3.228 3.557 -0.007 16.657 40

24

­50

510

BR

R e

stim

ates

1990 1995 2000 2005 2010Publication Year

95% CI Fitted valuesbrrest

Figure 1: The coe¢ cient estimates of the bene�t replacement rate in the unemploymentequations and the publication years.

­6­4

­20

24

BD

 est

imat

es

1990 1995 2000 2005 2010Publication Year

95% CI Fitted valuesbdest

Figure 2: The coe¢ cient estimates of the bene�t duration in the unemployment equationsand the publication years.

25

Table2:ThedescriptionofMeta-independentvariables

A)Data

Mean

Std.

#obs.

ubsum

1ifOECDsummarymeasureofunemploymentbene�tisused.0.279

0.449

390

bdyear

1ifbene�tdurationisexpressedinyears.

0.378

0.487

119

timinvinst

1iftimeinvariantinstitutionsdataareused.

0.083

0.276

409

avg5

1ifthedataare5-yearaverages.

0.328

0.470

409

lterm

1iflong-termunemploymentrateisanalyzed.

0.125

0.331

409

youthun

1ifyouthunemploymentrateisanalyzed.

0.027

0.162

409

austria

1ifAustriaisincludedinthesample.

0.878

0.328

409

ireland

1ifIrelandisincludedinthesample.

0.880

0.325

409

spain

1ifSpainisincludedinthesample.

0.741

0.439

409

portugal

1ifPortugalisincludedinthesample.

0.660

0.474

409

newzealand

1ifNewZealandisincludedinthesample.

0.663

0.473

409

switzerland

1ifSwitzerlandisincludedinthesample.

0.650

0.477

409

year6090

1ifthestudyusedthedatafrom

1960to1990s.

0.394

0.489

409

time60

1ifthetimeperiod1960sareused.

0.403

0.491

409

time70

1ifthetimeperiod1970sareused.

0.455

0.499

409

time80

1ifthetimeperiod1980sareused.

0.914

0.280

409

time90

1ifthetimeperiod1990sandaboveareused.

0.892

0.310

409

B)EstimationMethod

ols

1ifOLSisused.

0.337

0.473

409

re1ifrandom

e¤ectsmodelbyGLSisused

0.333

0.472

409

fe1if�xede¤ectsmodelisused

0.208

0.406

409

nonlinear

1ifnonlinearmodelisused.

0.078

0.269

409

26

C)ModelSpeci�cation

Mean

Std.

#obs.

�rstdi¤

1ifthe�rst-di¤erenceofunemploymentisestimated.

0.144

0.352

409

lagun

1ifauto-regressivemodelbyoneperiodisspeci�ed.

0.313

0.464

409

timdum

1iftimedummyisspeci�ed.

0.592

0.492

409

countrydum

1ifcountrydummyisspeci�ed.

0.560

0.497

409

tfps

1iftotalfactorproductivityshocksarecontrolled.

0.225

0.418

409

rinterest

1ifrealinterestrateiscontrolled.

0.276

0.448

409

in�a

1ifchangesinpriceleveliscontrolled.

0.279

0.449

409

home

1ifowneroccupationrateisspeci�ed.

0.037

0.188

409

lds

1iflabourdemandshockiscontrolled.

0.117

0.322

409

ud1ifuniondensityiscontrolled.

0.721

0.449

409

coord

1ifcoordinationinwagebargainingiscontrolled.

0.765

0.424

409

ep1ifemploymentprotectioniscontrolled.

0.736

0.441

409

uc1ifunioncoverageiscontrolled.

0.122

0.328

409

ttax

1iftotaltaxonlabouriscontrolled.

0.073

0.261

409

twedge

1iftaxwedgeiscontrolled.

0.469

0.500

409

D)Others

publication

theyearwhenthestudyispublished,non-binary

2002.335

4.451

409

rank

1ifpublishedinthejournalwiththeranking3orabove

0.230

0.421

409

27

Table 3: The Baseline Probit Estimation of the Binary Dependent Variables.BRR10 BD10

portugal 0.011 0.510(0.159) (0.228)*

avg5 -0.248 0.350(0.113)** (0.125)***

ubsum/bdyear 0.331 0.667(0.134)** (0.197)**

lterm -0.226 0.297(0.112)* (0.181)

youthun 0.376 dropped(0.143)*

timinvinst 0.511 0.431(0.102)** (0.177)

fe 0.088 0.886(0.133) (0.085)***

lds 0.332 -0.022(0.127)** (0.001)***

home -0.318 0.424(0.122)** (0.120)**

lagun -0.231 -0.816(0.101)** (0.262)

rank -0.349 -0.168(0.211) (0.395)

# of obs. 382 96pseudo-R2 0.346 0.502

Note: The numbers in the parentheses are robust standard errors for clustered samplesby paper. ***, **, and * denote signi�cance of 1%, 5%, and 10%, respectively. Theexplanatory variables year6090, �rstdi¤ , in�a, rinterest, tfps, ud, coord, twedge, and ttaxare not presented in the table. In the regression of BD10, variables year6090, youthun,�rstdi¤ , and coord are dropped because they predict the failure perfectly.

28

Table 4: The Baseline OLS Estimation.BRR_OLS BD_OLS BRRBD_OLS

portugal 0.118 0.390 1.091(0.223) (0.265) (0.054)***

avg5 -0.757 0.196 -8.911(0.425)* (0.174) (0.171)***

ubsum/bdyear -0.883 0.757 -14.354(0.349)** (0.309)** (0.673)***

lterm -0.643 0.010 dropped(0.355)* (0.193)

youthun -0.632 -0.364 dropped(0.316)* (0.527)

timinvinst -0.221 0.117 dropped(0.289) (0.482)

fe -0.030 1.667 1.411(0.293) (0.430)*** (0.327)***

lds 1.561 -1.460 -2.695(0.402)*** (1.049) (1.149)*

home -0.092 0.552 0.879(0.322) (0.247)** (0.396)*

lagun -0.478 -0.663 -10.427(0.478) (0.442) (0.054)*

rank -0.266 -0.146 -1.184(0.344) (0.270) (0.330)**

constant 1.385 0.068 11.759(0.521)** (0.766) (0.638)***

# of obs. 382 111 40F-stat. 24.25 . .R2 0.459 0.346 0.962

Note: The numbers in the parentheses are robust standard errors for clustered samplesby paper. ***, **, and * denote signi�cance of 1%, 5%, and 10%, respectively. Theexplanatory variables year6090, �rstdi¤ , in�a, rinterest, tfps, ud, coord, twedge, and ttaxare not presented in the table.

29

Table 5: The Sensitivity Analysis for Bene�t Replacement Rate.Probit (BRR) OLS (BRR) WLS (BRR)(1) (2) (3) (4) (5)

portugal 0.258 0.467 1.094(0.158) (0.438) (0.384)***

austria -0.530 0.650(0.097)*** (0.271)**

ireland -0.576 -0.962(0.084)*** (0.729)

newzeland 0.511 0.157(0.131)*** (0.438)

spain 0.360 -1.000(0.126)** (0.707)

switzerland -0.485 -0.936(0.143)*** (0.470)*

avg5 -0.272 -0.188 -0.662 -1.257 -1.693(0.100)*** (0.123) (0.321)** (0.695)* (0.318)***

ubsum 0.559 0.460 -0.452 -1.777 -2.971(0.088)*** (0.093)*** (0.256)* (0.679)** (0.590)***

year6090 -0.008 -0.681 -0.815(0.167) (0.397)* (0.426)*

time60 -0.009 0.241(0.236) (0.382)

time70 0.131 -0.834(0.163) (0.469)*

time80 0.082 -0.036(0.156) (0.362)

lterm -0.312 -0.457 -0.332 -1.879 -0.225(0.093)*** (0.093)*** (0.280) (0.826)** (0.711)

youthun 0.455 0.192 -0.028 -1.300 -2.386(0.107)** (0.229) (0.255) (0.709)* (0.743)***

timinvinst 0.425 0.101 -0.903 -0.653 -0.513(0.09)*** (0.218) (0.452)* (0.369)* (0.965)

fe 0.288 -0.259 -1.051(0.162)* (0.379) (0.162)***

re -0.102 0.018(0.121) (0.253)

ols 0.186 -0.060(0.142) (0.235)

nonlinear -0.025 0.373(0.161) (0.445)

30

�rstdi¤ -0.621 -0.529 -0.326 -0.106 0.057(0.055)*** (0.083)*** (0.187)* (0.557) (0.633)

in�a 0.055 -0.223 0.556 1.821 2.666(0.180) (0.165) (0.357) (0.791)** (0.646)***

rinterest -0.037 -0.242 0.154 -0.198 0.392(0.213) (0.176) (0.202) (0.223) (0.996)

lds 0.334 0.334 1.774 2.014 1.669(0.125)** (0.129)** (0.573)*** (0.650)*** (0.453)***

home -0.352 -0.381 -0.083 0.590 1.245(0.133)* (0.121)** (0.350) (0.668) (0.295)***

tfps 0.104 0.186 -0.580 -1.315 -0.327(0.148) (0.174) (0.429) (0.515)** (1.291)

ud 0.019 0.106 -0.888 -2.071 -0.615(0.133) (0.159) (0.373)** (0.806)** (0.426)

coord 0.394 0.361 0.660 0.266 0.091(0.104)*** (0.118)*** (0.305)** (0.267) (0.343)

twedge 0.053 0.131 0.671 0.783 1.726(0.126) (0.139) (0.306)** (0.419)* (0.504)***

ttax 0.443 0.535 0.498 -0.707 0.424(0.140)** (0.051)*** (0.484) (0.615) (0.513)

lagun -0.034 -0.040 -0.435 -0.844 -1.665(0.096) (0.114) (0.295) (0.911) (0.697)

ep -0.035(0.114)

uc -0.088(0.132)

bd -0.279(0.174)

brrbd 0.428(0.140)*

publication -0.083(0.017)***

constant 2.488 2.580 1.761(1.277)* (0.940)** (1.099)

# of obs. 382 380 382 175 352pseudo-R2 0.441 0.434 0.575 0.694 0.721

Note: The numbers in the parentheses are robust standard errors for clustered samples bypaper. ***, **, and * denote signi�cance of 1%, 5%, and 10%, respectively. In column (4),only the observations that are signi�cant at the 5% level are considered. Column (5) is theweighted least squares (WLS) estimation with weights equal to the inverse of the standarderrors of each observation.

31

Table6:TheSensitivityAnalysisforBene�tDurationandtheInteractionofBene�tReplacementRateandBene�t

Duration.

Probit(BD)

OLS(BD)

WLS(BD)OLS(BRRBD)WLS(BRRBD)

(1)

(2)

(3)

(4)

(5)

(6)

(7)

portugal

0.399

-0.182

0.432

1.107

0.741

(0.329)

(0.113)

(0.544)

(0.199)***

(0.309)*

ireland

0.305

0.044

(0.264)

(0.012)

spain

0.508

-0.538

(0.290)

(0.202)***

avg5

0.228

0.112

-0.371

0.202

-0.504

-8.812

-8.905

(0.199)

(0.036)***

(0.560)

(0.539)

(0.488)

(0.096)***

(0.062)***

year6090

-0.196

0.056

-2.617

1.339

(0.103)***

(0.223)

(2.318)

(0.603)*

time60

-0.219

-0.120

2.022

(0.099)**

(0.481)

(0.199)***

time70

-0.094

-1.015

-0.954

(0.275)

(0.598)

(1.847)

time80

0.456

dropped

(0.460)

time90

0.351

(0.128)

bdyear

0.315

10.213

0.088

0.410

-13.153

-15.673

(0.184)

(0.000)***

(0.545)

(0.329)

(0.452)

(1.005)***

(0.553)***

lterm

0.179

0.258

-0.035

0.083

0.040

dropped

dropped

(0.211)

(0.209)*

(0.312)

(0.333)

(0.303)

youthun

dropped

dropped

-0.418

dropped

1.418

dropped

dropped

(0.628)

(2.044)

�rstdi¤

dropped

dropped

-0.311

dropped

0.583

dropped

-1.129

(0.813)

(0.282)*

(0.945)

timinvinst

-0.646

-0.232

-0.491

-0.538

2.382

dropped

dropped

(0.193)

(0.066)***

(0.866)

(0.310)

(2.208)

fe1

1.737

2.658

1.592

0.525

(0.000)***

(0.392)***

(0.797)***

(0.433)***

(0.644)

re-0.937

-2.661

(0.048)***

(1.124)**

ols

-0.640

-1.492

(0.213)**

(0.912)

nonlinear

-0.228

-1.737

(0.577)

(1.003)*

32

in�a

-0.152

-1-1.002

-0.523

1.173

-4.183

-2.418

(0.321)

(0.000)***

(0.735)

(0.676)

(1.027)

(1.083)***

(1.014)*

rinterest

-0.739

11.277

-.0131

1.872

-3.114

-5.793

(0.136)***

(0.000)***

(1.323)

(0.193)

(2.474)

(0.790)***

(0.217)***

lds

0.957

-1-0.647

dropped

1.697

-4.578

-0.260

(0.016)***

(0.000)***

(2.254)

(1.697)

(1.480)**

(0.883)**

home

0.372

0.914

0.244

0.072

0.941

-0.185

0.960

(0.116)**

(0.046)***

(0.198)

(0.132)

(0.728)

(0.530)

(0.359)**

tfps

-0.619

10.618

dropped

-0.765

5.907

4.269

(0.009)***

(0.000)***

(1.962)

(2.143)

(0.963)***

(0.688)***

ud0.655

0.081

0.405

0.843

0.528

-0.426

0.445

(0.259)*

(0.047)***

(0.364)

(0.406)*

(0.654)

(0.350)

(0.175)**

coord

dropped

dropped

0.697

-1.042

0.875

dropped

dropped

(0.614)

(0.500)*

(1.872)

twedge

0.012

-10.113

-0.340

-1.422

0.199

-0.021

(0.277)

(0.000)***

(0.463)

(0.238)

(0.778)*

(0.199)

(0.309)

ttax

0.152

0.690

0.856

-0.012

-0.538

1.181

-0.699

(0.303)

(0.119)***

(0.548)

(0.605)

(0.949)

(0.554)

(0.492)

lagun

-0.390

-0.305

-0.540

-0.927

-0.915

-10.559

-10.777

(0.417)

(0.147)***

(0.688)

(0.605)

(0.894)

(0.199)***

(0.309)***

publication

-0.235

-0.043

-0.480

(0.097)***

(0.101)

(0.150)**

ep-0.016

dropped

(0.079)

uc1

dropped

(0.000)***

brr

0.252

1.358

-0.062

(0.075)***

(0.585)**

(0.616)

brrbd

1.000

-1.453

(0.002)***

(2.457)

constant

85.028

1.251

-0.892

973.300

14.074

(202.287)(0.371)***

(1.722)

(301.649)

(1.8314)***

#obs.

9696

111

50111

4040

(Pseudo)-R2

0.430

0.724

0.485

0.702

0.345

0.958

0.977

Note:Thenumbersintheparenthesesarerobuststandarderrorsforclusteredsamplesbypaper.***,**,and*denotesigni�cance

of1%,5%,and10%,respectively.Incolumn(4),onlytheobservationsthataresigni�cantatthe5%

levelareconsidered.

33

Table7:SummaryoftheCharacteristicsofthePapersIncludedintheMeta-Analysis

ID.

Paper

#Countries

Period

Data

Est.method

1Nickell(1997)

20,19

1983-1994

6-yravg

RE

2Nickell(1998)

20,19

1983-1994

6-yravg

RE

3Nickelletal(2001)

15,20,19

1960-95

yearly

RE

4Nunziata(2002)

18-20

1960-95,1970-95

yearly

FE,OLS

5Nickelletal(2005)

20,19

1961-95

yearly

RE,NLS

6Blanchard&Wolfers(2000)

201960-95

5-yravg

NLS

7Elmeskovetal(1998)

18,19

1983-95

yearly

RE

8Belot&vanOurs(2001)

181960-94

5-yravg

FE,OLS

9Belot&vanOurs(2004)

171960-99

5-yravg

FE,OLS

10Scarpetta(1996)

171983-93

yearly

FGLS

11Bertolaetal(2001)

201960-96

5-yravg

OLS,NLS

12Baccaro&Rei(2005)

181960-98

yearly,5-yravg

RE,OLS,NLS

13Bakeretal(2003)

20,19

1985-94,1960-99

5-yravg

RE

14IMF(2003)

201960-98

yearly

RE

15OECD(1999)

191985-90,1992-97

6-yravg

RE

16Esping-Andersen&Regini(2000)

201996

1year

OLS

17Amableetal(2006)

181980-2004

yearly

OLS,FE,RE

18Gri¢thetal(2006)

141986-2000

yearly

OLS,IV

19Fitoussi(2000)

191980s-1990s

1period

OLS

20Jackmanetal(1996)

20,19

1983-94

6-yravg

OLS

21Daveri&Tabellini(2000)

141965-95

5-yravg

OLS

22Chenetal(2003)

191960-99

10-yravg

OLS

23Jackmanetal(1990)

141971-88

yearly

IV24

Bakeretal(2004)

201960-98

yearly,5-yravg

RE

25Bassanini&Duval(2006)

201982-2003

yearly,5-yravg

FE,RE,OLS,NLS

26Boone&vanOurs(2004)

201985-99

yearly,5-yravg

RE,FE

27Macculloch&DiTella(2005)

211984-90

yearly

RE,LSVDV,GMM,FE

28Alganetal(2002)

171960-2000

5-yravg

OLS,GLS

29Burda(1988)

111985,1979

1year

OLS

30Howell(2003)

201989-94,1995,2001

5-yravg,1period

OLS

31Kenworthy(2002)

161980-97

yearly

OLS

32Amableetal(2007)

181980-94

yearly

RE,FE,OLS

33Garibaldi&Violante(2005)

171960-2000

5-yravg

FE

34Addison&Teixeira(2005)

201956-99

6-yravg,5-yravg

RE,NLS

34

Table8:SummaryoftheMeta-ObservationsforBene�tReplacementRatioandBene�tDuration

BRR

BRR10

BD

BD10

Id.

Paper

Mean

Min

Max

#obs.

Mean

Mean

Min

Max

#obs.

Mean

1Nickell(1997)

0.011

0.011

0.011

31

0.127

0.043

0.25

30.667

2Nickell(1998)

0.012

0.011

0.013

30.667

0.132

0.045

0.25

30667

3Nickelletal(2001)

1.48

0.3

2.21

30.667

0.363

0.22

0.47

31

4Nunziata(2002)

2.548

0.193

4.356

201

0.901

0.006

3.228

200.85

5Nickelletal(2005)

2.124

1.88

2.4

51

0.39

0.34

0.47

50.6

6Blanchard&Wolfers(2000)

0.096

0.006

0.025

110.818

0.212

0.157

0.267

71

7Elmeskovetal(1998)

0.096

0.08

0.11

71

..

.0

.8

Belot&vanOurs(2001)

-0.617

-2.14

1.09

70.286

..

.0

.9

Belot&vanOurs(2004)

-0.074

-0.24

0.06

70.286

..

.0

.10

Scarpetta(1996)

0.109

0.02

0.18

300.833

..

.0

.11

Bertolaetal(2001)

0.001

0.001

0.001

21

..

.0

.12

Baccaro&Rei(2005)

-0.007

-0.03

0.021

630.048

-0.485

-1.433

0.225

90

13Bakeretal(2003)

-0.108

-0.61

0.064

50

-2.628

-6.685

3.955

50

14IMF(2003)

-0.008

-0.044

0.012

40.25

..

.0

.15

OECD(1999)

0.011

0.01

0.02

70.571

00

07

016

Esping-Andersen&Regini(2000)

0.004

-0.003

0.012

30

0.077

-0.02

0.138

30.667

17Amableetal(2006)

1.463

0.464

3.580

330.364

..

.0

.18

Gri¢thetal(2006)

7.800

2.36

10.72

70.857

..

.0

.19

Fitoussi(2000)

-0.023

-0.2

0.12

30.667

0.533

0.2

0.79

30.667

20Jackmanetal(1996)

0.008

0.004

0.011

30.333

0.097

0.04

0.16

30.333

21Daveri&Tabellini(2000)

0.109

0.05

0.18

140.643

..

.0

.22

Chenetal(2003)

0.05

0.01

0.12

60.5

0.203

-0.5

1.3

60.333

23Jackmanetal(1990)

0.435

0.28

0.67

40.5

0.165

0.14

0.22

41

24Bakeretal(2004)

0.007

-0.019

0.034

100.4

-1.340

-2.197

-.482

20

25Bassanini&Duval(2006)

0.106

0.04

0.21

401

2.64

2.64

2.64

11

26Boone&vanOurs(2004)

0.242

0.083

0.512

61

..

.0

.27

Macculloch&DiTella(2005)

-0.021

-0.555

0.202

170

..

.0

.28

Alganetal(2002)

0.006

0.005

0.007

40.5

0.002

-0.004

0.005

40

29Burda(1988)

0.065

0.03

0.14

160.938

..

.0

.30

Howell(2003)

0.053

-0.011

0.11

100.2

1.108

0.54

1.53

100.9

31Kenworthy(2002)

..

.0

.0.09

0.04

0.19

50.2

32Amableetal(2007)

1.559

-0.83

3.129

130.615

..

.0

.33

Garibaldi&Violante(2005)

-0.03

-0.03

-0.03

40

..

.0

.34

Addison&Teixeira(2005)

0.001

-0.031

0.036

120.417

0.005

-0.016

0.032

80.25

Note:1.Austriaexcludedforthelong-,respective,theshort-term

unemploymentrates.2.Portugalexcluded.

35

Appendix B Meta-Analysis References

1. Nickell, S. (1997), "Unemployment and Labor Market Rigidities: Europe versusNorth America," Journal of Economic Perspectives, 11 (3), 55-74.

2. Nickell, S. (1998), "Unemployment: Questions and Some Answers," The Eco-nomic Journal, 108 (448), 802�816.

3. Nickell, S., L. Nunziata, W. Ochel, and G. Quintini (2001), "The BeveridgeCurve, Unemployment and Wages in the OECD from the 1960s to the 1990s,"CEP Discussion Papers 0502.

4. Nunziata, L. (2002), "Unemployment, Labour Market Institutions and Shocks,"Economics Papers 2002-W16, Economics Group, Nu¢ eld College, University ofOxford.

5. Nickell, S., L. Nunziata, and W. Ochel: "Unemployment in the OECD Sincethe 1960s. What Do We Know?," Economic Journal, 115 (500), 1-27.

6. Blanchard, O. and J. Wolfers (2000), "The Role of Shocks and Institutions inthe Rise of European Unemployment: the Aggregate Evidence," The EconomicJournal, 110 (462), 1�33.

7. Elmeskov, J., J. P. Martin, and S. Scarpetta (1998), "Key Lessons For LabourMarket Reforms: Evidence From OECD Countries�Experience," Swedish Eco-nomic Policy Review, 5 (2), 205-258.

8. Belot, M. and J. C. van Ours (2001), "Unemployment and Labor Market In-stitutions: An Empirical Analysis," Journal of the Japanese and InternationalEconomies, 15 (4), 403-418.

9. Belot, M. and J. C. van Ours (2004), "Does the recent success of some OECDcountries in lowering their unemployment rates lie in the clever design of theirlabor market reforms?," Oxford Economic Papers, 56 (4), 621-642.

10. Scarpetta, S. (1996), "Assessing the Role of Labour-Market Policies and Insti-tutional Factors on Unemployment: A Cross-Country Study�, OECD EconomicStudies No. 26, Paris.

11. Bertola, G., F. D. Blau, and L. M. Kahn (2001), "Comparative Analysis ofLabor Market Outcomes: Lessons for the Us from International Long-Run Ev-idence," NBER Working Paper 8526.

12. Baccaro, L. and D. Rei (2005), "Institutional Determinants of Unemployment inOECDCountries: A Time Series Cross-Section Analysis (1960-98)," DP/160/2005,International Institute for Labour Study, Geneva.

36

13. Baker, D., A. Glyn, D. Howell, and J. Schmitt (2003), "Labor Market Insti-tutions and Unemployment: A Critical Assessment of the Cross-Country Evi-dence," CEPA Working Paper 2002-17, Center for Economic Policy Analysis,New School University.

14. Chapter 4, "Unemployment and Labor Market Institutions: Why Reform Payo¤,"World Economic Outlook (2003), World Economic and Financial Surveys:Growth and Institutions, International Monetary Fund.

15. Chapter 2, "Employment Protection and Labour Market Performance," OECDEmployment Outlook (1999), Paris.

16. Esping-Andersen, G. and M. Regini (2000), "Who is Harmed by Labour Mar-ket Regulations? Quantitative Evidence," Chapter 3, Why Deregulate LabourMarkets?, Oxford University Press.

17. Amable, B., L. Demmou, and D. Gatti (2006), "Institutions, Unemploymentand Inactivity in the OECD countries," Working Paper N� 2006 - 16, Paris-Jourdan Sciences Economiques.

18. Gri¢ th R., R. Harrison, and G. Macartney (2006), "Product Market Reforms,Labour Market Institutions and Unemployment," IFS Working Papers, W06/06,The Institute for Fiscal Studies.

19. Fitoussi, J-P., D. Jestaz, E. S. Phelps, and G. Zoega (2000), "Roots of theRecent Recoveries : Labor Reforms or Private-Sector Forces?" Brookings Paperson Economic Activity, 2000 (1), 237-311.

20. Jackman R., R. Layard, and S. Nickell (1996), "Combating Unemployment:Is Flexibility Enough?" CEP-Discussion Paper No. 293, Centre for EconomicPerformance.

21. Daveri, F. and G. Tabellini (2000), "Unemployment, Growth and Taxation inIndustrial Countries," Economic Policy, 15 (3), 49-90.

22. Chen, Y-F., D. Snower, and G. Zoega (2002), "Labour Market Institutions andMacroeconomic Shocks," IZA Discussion Papers 539, Institute for the Studyof Labor (IZA).

23. Jackman, R., C. Pissarides, and S. Savouri (1990), �Labour Market Policies andUnemployment in the OECD,�Economic Policy 11, October.

24. Baker, D., A. Glyn, D. Howell, and J. Schmitt (2004), "Unemployment andLabor Market Institutions: the Failure of the Empirical Case of Deregulation,"CEPA Working Paper 2004-4, Center for Economic Policy Analysis, New ShoolUniversity.

37

25. Bassanini, A. and R. Duval (2006), "Employment Pattern in OECD Coun-tries: Reassessing the Role of Policies and Institutions," DELSA/ELSA/WDSEM(2006)4, OECD Social, Employment, and Migration Working Papers, No.35.

26. Boone, J. and J. C. van Ours (2004), "E¤ective Active Labor Market Policies,"IZA Discussion Papers No.1335, Institute for the Study of Labor (IZA).

27. Di Tella, R. and R. MacCulloch (2005), "The Consequences of Labor Mar-ket Flexibility: Panel Evidence Based on Survey Data," European EconomicReview, 49 (5), 1225-1259.

28. Algan Y., P. Cahuc, and A. Zylberberg (2002), "Public Employment andL-labour Market Performance," Economic Policy, 17 (34), 7-66.

29. Burda, M. (1988), "Wait Unemployment in Europe," Economic Policy, 3 (7),391-416.

30. Howell, D. R. (2003), "The Micro-Foundations of High Unemployment in De-veloped Countries: Are Labor Market Rigidities the Problem?," Chapter 7 inHarris and Goodwin, eds. New Thinking in Macroeconomics: Social, Institu-tional, and Environmental Perspectives (Elgar, 2003).

31. Kenworthy, L. (2002), "Corporatism and Unemployment in the 1980s and 1990s,"American Sociological Review, 67 (3), 367-388.

32. Amable, B., L. Demmou, and D. Gatti (2007), "Employment Performance andInstitutions: New Answers to an Old Question," Discussion Paper No.2731,Institute for the Study of Labor (IZA).

33. Garibaldi, P. and G. L. Violante (2005), "The Employment E¤ects of SeverancePayments with Wage Rigidities," The Economic Journal, 115 (506), 799-832.

34. Addison, J. T. and P. Teixeira (2005), "What Have We Learned about theEmployment E¤ects of Severance Pay? Further Iterations of Lazear et al.,"Empirica, 32 (2), 345-368.

38