An Empirical Analysis of the Dynamics of the Welfare State: The Case of Benefit Morale -super-*

Abstract

Does the supply of a welfare state create its own demand? Many economic scholars studying welfare arrangements refer to Say's law and insinuate a self-destructive welfare state. However, little is known about the empirical validity of these assumptions and hypotheses. We study the dynamic effect of different welfare arrangements on benefit fraud. In particular, we analyze the impact of the welfare state on the respective social norm, i.e. benefit morale. It turns out that a high level of public social expenditures and a high unemployment rate are associated with a small positive (or no) immediate impact on benefit morale, which however is (partly) crowded out by adverse medium and long run effects. In contrast to earlier studies we do not find that younger birth cohorts have lower values of benefit morale. Copyright © 2010 Blackwell Publishing Ltd.

Full text

Page 1

An Empirical Analysis of the Dynamics of the Welfare State:
The Case of Benefit Morale?
Martin Halla
University of Linz & IZA
Mario Lackner
University of Linz
Friedrich G. Schneider
University of Linz, IZA, CESifo & CREMA
1 INTRODUCTION
Economic scholars studying the welfare state often discuss disincentive effects
of welfare arrangements, and of the taxes to finance them, on economic
behavior, which may create new problems (Lindbeck, 1995a,b; Gouyette and
Pestieau, 1999; Ravallion, 2001; Lindbeck, Nyberg and Weibull, 2003;
HenreksonandPersson,2004;BeaulierandCaplan,2007).Thebasicdilemma
ofthewelfarestateisattributedtothedirecteffectofmoregenerousbenefitson
thenumberofrecipients.Thiseffectoperatesthroughthreedifferentchannels:
taxdistortions,moralhazardandbenefitfraud.1Theliteraturepointsoutthat
KYKLOS, Vol. 63 – February 2010 – No. 1, 55–74
r 2010 Blackwell Publishing Ltd., 9600 Garsington Road,
Oxford OX4 2DQ, UK and 350 Main Street, Malden, MA 02148, USA
55
?
Corresponding author: Mario Lackner, Johannes Kepler University of Linz, Department of
Economics, Altenbergerstr. 69, 4040 Linz, ph.: 143 70 2468 8246, fax: 143 70 2468 28246, email:
mario.lackner@jku.at. All estimations discussed in Section 3 are based on data generously provided
by Friedrich Heinemann. For helpful discussions and comments we would like to thank Eddy
Bekkers, Johann K. Brunner, Joseph F. Francois, Franz Hackl, Friedrich Heinemann, Karin Mayr,
Gerald Pruckner, Johann Scharler, two anonymous referees and the editors. The usual disclaimer
applies. This paper was partly written during Martin Halla’s visiting scholarship at the Center for
Labor Economics at the University of California, Berkeley. He would like to give thanks for the
stimulating academic environment and hospitality there. Financial support from the Austrian FWF
(NFN Labor Economics and the Welfare State) is gratefully acknowledged.
1Benefit fraud seem to be particularly pervasive in the case of social benefits, sick benefits, and
unemployment benefits. Greenberg, Moffitt and Friedman (1981); Greenberg and Halsey (1983) study
transferprograms(negative-incometaxplans)intheU.S.andfindsubstantialincomeunderreportingfor
up to 50 percent of certain subgroups of the population. Skogman Thoursie (2004) provides empirical

Page 2

it is necessary to discuss the disincentive effect of welfare arrangements in the
context of a dynamic interaction between market behavior and political
behavior over time. It is hypothesized that individuals do not respond to
changesinsucheconomicincentivesimmediately,sincetheyareconstrainedby
social norms (Elster, 1989) for some time. Therefore, the disincentive effects
may materialize only with considerable time lags.
Thetheoreticalliteraturerestsupontestableassumptionsandoffersconcrete
hypotheses on welfare-state dynamics involving endogenous changes in social
norms and political preferences. However, little is known about the empirical
validity of these assumptions and hypotheses. The empirical research in this
areaisstillatanearlystage.Lindbeck(2003)statesthat‘theoryandspeculation
are far ahead of systematic empirical research in the field of welfare-state
dynamics’. This fact is surprising since there is a widespread concern about
abuse and dishonesty in social welfare programs. Heinemann (2008) (hence-
forthFH)isthefirstattempttoempiricallytestthehypothesesonthedynamic
relationofthewelfarestateandsocialnormsprovidedbytheory.2Hisanalysis
of benefit morale (i. e. the individual reluctance to exploit the welfare state via
benefit fraud) puts forward that an increase in social benefits and in the
unemployment rate over the preceding twenty years is associated with
substantial decrease in benefit morale. Moreover, he claims that later birth
cohorts have significantly lower levels of benefit morale whereas age has no
effect.
In this paper we try to extend and deepen the analysis of FH. We start with
his three basic hypotheses and highlight several methodological problems to
overcome in order to test these hypotheses. For instance, we demonstrate the
great importance of controlling for country fixed-effects for this type of
analysis. The generosity of the welfare state is correlated with unobserved
country-specific time-invariant heterogeneity in a way that disregarding
country fixed-effects can diametrically reverse results. In a second step we will
suggest a refined econometric framework which is used to present our own
empirical results based on an extended data set. Our empirical model allows,
among others, public social expenditures and the level of unemployment to
evidence for abuse of the sickness insurance system in Sweden by comparing the change between the
numberofmenandwomenwhoreportsickduringpopularsportingevents.Thedifference-in-differences
approach shows that the number of men who reported sick increases considerably in order to watch the
sporting event on television. Kingston, Burgess and St. Louis (1986) identify frequent overpayments in
unemployment insurance systems of five U.S. states, indicating that many claimants falsely certify that
theyhaveactivelysoughtajob.WolfandGreenberg(1986)findfraudratesof2to4percentintheAidto
Families with Dependent Children and Food Stamps entitlement programs.
2Algan and Cahuc (2008) study the impact of benefit morale on the design of public insurance policies
withinlabormarkets.Theyprovideempiricalevidencethatcountriesdisplayinghighbenefitmoraletend
toinsuretheirworkersthroughunemploymentbenefitsinsteadofusingstringentemploymentprotection.
56
r 2010 Blackwell Publishing Ltd.
MARTIN HALLA/MARIO LACKNER/FRIEDRICH G. SCHNEIDER

Page 3

havevaryingeffectsonbenefitmoraleintheshort,mediumandlongrun.This
more flexible modeling reveals interesting dynamic effects of welfare arrange-
ments onbenefit morale.A highlevel of public social expendituresand a large
number of unemployed in the current period have small positive (or no) short
run effects on benefit morale, however, these are (partly) crowded out by
adverse medium and long run effects. This finding is consistent with the
hypothesisthatdisincentiveeffectsofagenerouswelfarestatematerializeonly
with some time lag. Further, in sharp contrast to FH we can not support the
hypothesis of the degeneration of younger cohorts’ benefit morale. Once the
problem of linear dependency among age, period and cohort is adequately
addressed this effect vanishes.
The paper is organized as follows: In Section 2 we briefly summarize the
hypotheses on the dynamic effects of the welfare state put forward by FH.
Section 3 discusses several methodological problems to overcome in order to
testthesehypotheses.OurrefinedestimationstrategyispresentedinSection4,
and Section 5 discusses our new results. Section 6 concludes the paper.
2 HYPOTHESES ON WELFARE STATE DYNAMICS AND
BENEFIT MORALE
Does the supply of a welfare state create its own demand? Many economic
scholars studying welfare arrangements refer to Say’s law and insinuate a
self-destructive welfare state. The theoretical literature (e. g. Lindbeck,
1995a,b, 2003) emphasizes that it is necessary to account for social norms
and to discuss the consequences of welfare state arrangements in a dynamic
context. This should reflect the interacting adjustments in basic behavior
patterns of households, firms, interest-groups, politicians and publicsector
administrators over time. The fundamental supposition is that individuals
donotrespondtochangesineconomicincentivesimmediately,sincetheyare
constrained by social norms for some time. Individuals are assumed to
experience disutility when violating social norms. This can be explained by
intrinsic factors, a subjectively felt resistance to violate social norms, or
extrinsic factors such as a loss of reputation possibly accompanied by
punishment. Put differently, new incentives created by the welfare state are
in conflict with existing social norms. However, as time evolves individuals
gradually stop obeying initially existing norms and disincentive effects may
materialize with some time lags.
FH considers a particular channel, namely the role of benefit fraud in
the hypothesis of hazardous welfare state dynamics. In particular, he
empirically analyzes the effects of the extent of the welfare state on the
social norm with respect to benefit fraud. To measure the individual
r 2010 Blackwell Publishing Ltd.
57
THE CASE OF BENEFIT MORALE

Page 4

reluctance to exploit the welfare state via benefit fraud (i. e. benefit morale)
he employs large scale international survey data from the European and
World Values Survey.
Hypothesis 1 In order to test the fundamental supposition of Lindbeck
(1995a,b) – suggesting that generous welfare payments in preceding periods
reduce benefit morale today – FH first employs the change in social spending
over the preceding twenty years derived from the OECD Economic Outlook
Database. Firstly, we suggest to quantify the generosity of the welfare state by
public social expenditures measured as a percentage of GDP derived from the
OECDSocialExpenditureDatabase(OECD,2007).Thisdatabasesummarizes
information on aggregated public social expenditure grouped along nine core
socialpolicyareasandallowsustomeasurethelevelofaggregatepublicsocial
expenditure on an internationally comparable base. Secondly, and more
importantly, we propose a more flexible empirical model to capture the
dynamics suggested by the theoretical literature. In order to allow for varying
effects of a generous welfare state on benefit morale in the short, medium and
long run we consider the effect of public social expenditure in the current
period, with a lag of five years and with a lag of ten years. This specification
shouldallowustotestwhetherdisincentiveeffectsindeedmaterializeonlywith
some time lag.
Hypothesis2Ina second stepFHanalyzesthechangeintheunemployment
rate over the preceding twenty years. This hypothesis can be derived from
Lindbeck, Nyberg and Weibull (1999) who constitute an additional explana-
tion why disincentive effects of generous welfare arrangements are likely to be
stronger in the long than in the short run. They hypothesize that the level of
benefitmoraletendstofallwiththenumberofindividualslivingonbenefits,in
the sense that the guilt or shame connected with breaking the social norms is
thenreduced.Thoughtheunemploymentrateisnottheonlypossiblechoiceto
quantify the population share living on benefits, we follow FH since the
unemployment rate distinguishes itself from other measures by its wide
availability and comparability across countries. Again, we will focus on the
dynamic effects of unemployment.
Hypothesis3Finally,FHtestswhetherindividualswhoarebornatdifferent
stagesofthewelfarestateexhibitdifferentinitiallevelsofbenefitmorale.Since
thewelfarestateinallOECDmembercountrieshasbeenexpandingonaverage
overthelastdecades,hehypothesizesthatyoungerbirthcohortsshouldhavea
lower level of benefit morale. To test this hypothesis empirically, he simply
includes both the year of birth and the age of the respondent as explanatory
variables. As we will show, in order to test Hypothesis 3 accurately one has to
disentangle different types of time-related variation in benefit morale: the
effects of age, period, and birth cohort. After a thorough discussion of the
identification problem created by the exact linear dependency among age,
58
r 2010 Blackwell Publishing Ltd.
MARTIN HALLA/MARIO LACKNER/FRIEDRICH G. SCHNEIDER

Page 5

period and cohort (Fienberg and Mason, 1985) we will use a cross-classified
fixed-effectsmodel(YangandLand,2006)tocleanlyidentifytheeffectofbirth
cohorts.
3 METHODOLOGICAL ISSUES
In this section we discuss several methodological challenges to overcome in
ordertotestHypotheses1to3.Tohighlighttherelativeimportanceofdifferent
econometricissuesweanalyzetherobustnessofthefindingsofFH.InSection4
we will suggest a refined econometric framework which is used to present our
new set of results in Section 5.
3.1 Measurement of benefit morale
To our best knowledge there are two alternative data sources to study the
phenomenon of benefit morale, the European and World Values Surveys
(E/WVS) and the International Social Survey Programme. In order to guaran-
tee comparability with previous papers we employ data from the E/WVS
throughoutthepaper.Thesesurveysprovidedatafromrepresentativenational
samples (based on face-to-face interviews) of more than 80 countries. It
contains information on basic attitudes, beliefs and human values covering
religion, morality, politics, work and leisure. To date, four waves have been
conducted. The exact question on benefit morale in the E/WVS questionnaire
readsasfollows:‘Pleasetellmeforeachofthefollowingstatementswhetheryou
think it can always be justified, never be justified, or something in between:
Claiming governments benefits to which you are not entitled.’ Respondents are
asked to evaluate on an ordered scale from ‘never justifiable’ (1) to ‘always
justifiable’ (10). We use this question to construct our measure of benefit
morale.
The original question in the E/WVS questionnaire gives a measure of
benefit morale on a ten-point scale. FH and Algan and Cahuc (2008) use a
re-scaled version of this measure as a dependent variable, and as a
explanatory variable, respectively in their main analysis. They create a binary
measure which takes on the value one if the respondent answers ‘never
justifiable’ in the E/WVS, and zero otherwise. The authors argue that the
re-scaling should ease the interpretation of results. However, this mapping
from the ordered scale into a binary variable seems arbitrary. This specific
mapping groups together respondents with the highest level of benefit
morale and all other respondents. That means, this binary measure does not
capture a variationfrom thelowest level of benefit morale (the answer‘always
r 2010 Blackwell Publishing Ltd.
59
THE CASE OF BENEFIT MORALE

Page 6

justifiable’) to considerably higher level of benefit morale such as the answer
‘nine’.
Wethink that themost innocuousprocedureis to stick to theoriginal ten-
point scale provided by the E/WVS questionnaire but to reverse the scaling
suchthatahighervalueofthevariablealsoindicatesahigherlevelofbenefit
morale.3This measure makes use of the whole variation and has a straight
forward interpretation. In order to explore the sensitivity of the results with
respecttomappingfromtheorderedscaleintoabinaryvariablewecompare
the results for both variables based on the preferred specifications of FH.4If
we switch from the binary measure of benefit morale to the ordinal variable
no qualitative changes can be observed with respect to Hypothesis 1 to 3.
However, the control variable income turns out to be a statistically
significant determinant of benefit morale (detailed output available upon
request). Although there are no further substantial differences between the
results obtained from the two different versions, we cannot find any
argument in favor of reducing the full ten-point measure of benefit morale
which is available in the underlying data.
3.2 Time-related variation in benefit morale
In order to test Hypothesis 3 thoroughly one has to disentangle the effects of
age, period, and birth cohorts. Age effects represent variation associated with
different age groups brought about by experience of life (e.g. accumulation of
social capital), physiological changes and/or role or status changes. Period
effects are defined as variation over time periods that affect all age groups
simultaneously. These may result from from shifts in political, social or
economicalenvironments.Birthcohorteffectsrepresentchangesacrossgroups
of individuals born in the same year (Ryder, 1965). A birth cohort moves
throughlifetogetherandencountersthesamehistoricalandsocialeventsatthe
same ages.
Sincethesumofaperson’sageandhisyearofbirthisequaltothecurrent
year (i. e. the year of the survey), there exists an exact linear relationship
between these variables. Put differently, one of the variables is an exact
linear combination of the two others, and a general linear model cannot be
used to identify all three effects. An extensive literature across disciplines
discusses how to identify these three effects with different types of data and
in different settings (see, e. g., Glenn, 1976; Mason, Mason and Winsbor-
3The reversed scaling does only change the sign of the coefficient in a regression analysis but leaves the
absolute value unchanged.
4The correlation between the two measurements of benefit morale is about 0.74.
60
r 2010 Blackwell Publishing Ltd.
MARTIN HALLA/MARIO LACKNER/FRIEDRICH G. SCHNEIDER

Page 7

ough, 1976; Fienberg and Mason, 1979, 1985; Robertson, Gandini and
Boyle, 1999).
The E/WVS collects information onboth age and year of birth. The year of
the survey should be equal to the sum of the respondent’s age and his year of
birth. Unfortunately, there are many cases of age and/or year of birth
misreporting in the data.5A border line case is given by observations where
the sum of the respondent’s age and his year of birth is by one lower or higher
thanthesurveyyear.Thefirstcasemayoccuriftheinterviewtookplacebefore
the respondent’s birthday in the given year. The second case may occur if the
survey lasted until the beginning of the next year and the interview took place
aftertherespondent’sbirthday.6Sincetheavailablesurveydatadonotprovide
information on the month of birth or on the month of the interview, it is not
possibletodefinitelydistinguishbetweenobservationsthatnaturallydeviateby
one and corrupted observations. However, an auxiliary regression analysis
providesevidencethatthatthemajorityofthesedeviationsisduetothetiming
ofinterviewandthebirthdayratherthanerrorsinthedata.7Therefore,wekept
these observations throughout all our estimations, however, disregarded all
observations with strictly impossible combinations of respondents’ age and
year of birth.
FHneitheraddressestheobservationswithstrictlyimpossiblecombinations
of respondents’ age and year of birth, nor the identification problem of age,
period and birth cohort effects. His identification of birth cohort effects rests
apparently upon variation from (i) strictly impossible combinations of
respondents’ age and year of birth, (ii) the specific timing of interview and
birthday, and (iii) the omission of period effects.8
In columns (Ia) and (Ib) in Table 1 we augment FH’s two preferred
specifications by period fixed-effects and we see that the statistically significant
negativeimpactoftheyearofbirth(i.e.thebirthcohorteffect)disappearsinboth
specifications.Inanycase,thisresulthastobeinterpretedwithcautiontoo,since
5This type of misreporting has several potential sources. Mason and Cope (1987) mention (i) ignorance
about age, (ii) distortion of age to meet preconceptions about the relationship of age to other
characteristics or events, (iii) communication problems between interviewers and respondent, and (iv)
errors of recording or processing.
6In these cases, the survey year is assumed to be the preceding year, since the majority of interviews took
place then.
7A simply estimation, regressing a binary variable – taking on the value 1 if the the deviation from age,
periodandcohortsiseitherzeroorminusone–onthemonthwhenthefieldworkhasstarted/ended(which
variesacrosscountriesandyears)showsthattheearlierfieldworkhasstartedthemorelikelytherewillbea
deviation from a perfect linear relation between age, period and cohort. Note, only for the interviews
conducted in the year 1999 the month ofthe interview is available. In all other cases only information on
the month when the field work has started and ended is provided.
8It should be noted that FH controls in some specifications for period fixed-effects, however, not in the
specification on which his results on birth cohort effects are based.
r 2010 Blackwell Publishing Ltd.
61
THE CASE OF BENEFIT MORALE

Page 8

Table1
Determinants of benefit morale: methodological challenges to overcome in order to test
hypotheses 1 to 3.a
(Ia) (IIa)(IIIa) (Ib)(IIb) (IIIb)
HYPOTHESIS 1
D Social security benefitsb2 0.072??
0.028
(0.028)
2 0.003
(0.014) (0.034)
HYPOTHESIS 2
D Unemploymentc
2 0.053??
(0.019)
0.011
(0.028)
2 0.012
(0.012)
HYPOTHESIS 3
Year of birth 0.0003
(0.081)
2 0.082
(0.067)
0.011
(0.037)
0.023
(0.064)
2 0.067
(0.066)
0.006
(0.037)
SOCIO-ECONOMIC CONTROL VARIABLES
Age 0.019
(0.081)
0.139???
(0.027)
0.186???
(0.030)
0.025?
(0.012)
2 0.065
(0.066)
0.126???
(0.022)
0.188???
(0.026)
0.026??
(0.010)
0.028
(0.037)
0.128???
(0.019)
0.186???
(0.024)
0.035???
(0.008)
0.005
(0.003)
0.040
(0.063)
0.140???
(0.028)
0.186???
(0.026)
0.024??
(0.012)
2 0.051
(0.066)
0.131???
(0.023)
0.191???
(0.021)
0.028???
(0.008)
0.023
(0.037)
0.133???
(0.019)
0.182???
(0.022)
0.038???
(0.006)
0.006?
(0.003)
Female
Married
Income
School leaving age
Employed0.169???
(0.048)
0.110???
(0.038)
0.129???
(0.044)
0.092??
(0.036)
Unemployedd
2 0.314???
(0.084)
2 0.014
(0.047)
2 0.058??
(0.027)
2 0.084??
(0.035)
0.002
(0.008)
2 0.295???
(0.079)
2 0.022
(0.038)
2 0.053?
(0.026)
2 0.071??
(0.029)
2 0.005
(0.009)
Self-employedd
Out of labor forced
Size of town
Number of children
ATTITUDINAL CONTROL VARIABLES
Religious 0.023
(0.016)
0.107???
(0.026)
0.127???
(0.032)
Yes
No
8.496
(160.033)
0.025???
(0.006)
0.068???
(0.018)
0.132???
(0.029)
Yes
Yes
171.315
(132.091)
0.017??
(0.006)
0.056??
(0.020)
0.117???
(0.031)
Yes
Yes
2 13.730
(73.990)
0.022
(0.019)
0.095???
(0.022)
0.127???
(0.033)
Yes
No
2 36.531
(126.382)
0.028???
(0.007)
0.057???
(0.017)
0.136???
(0.032)
Yes
Yes
141.671
(130.985)
0.019???
(0.006)
0.046??
(0.017)
0.120???
(0.034)
Yes
Yes
2 3.231
(73.593)
Confidence in parliament
Patriotism
Period fixed-effectse
Country fixed-effects
Constant
No. of observations
R-squared
53,303
0.058
53,303
0.099
44,517
0.103
62,862
0.053
62,862
0.101
50,400
0.101
aAll data sources can be found in FH and/or in the Data appendixin Halla, Lackner and Schneider
(2009).Thedependentvariableisequaltobenefitmoralemeasuredonaten-pointscale,wherehigher
values indicate a higher level of benefit morale. Estimated using ordinary least squares.
?,??and???indicatestatisticalsignificanceatthe10-percentlevel,5-percentlevel,and1-percentlevel,
respectively.
b20 Year growth rate of social security benefits as percentage of GDP.
c20 Year growth rate of unemployment rate.
dBase group is employed.
ePeriodfixed-effectsareincludedasyeardummiesbetween1981and2000withtheyear1981asthebase.
62
r 2010 Blackwell Publishing Ltd.
MARTIN HALLA/MARIO LACKNER/FRIEDRICH G. SCHNEIDER

Page 9

the identification is solely due to the small deviations resulting from the specific
timing of interview and birthday. In Section 4, we will employ a cross-classified
fixed-effects model developed for micro-level data in the form of repeated
cross-sections (Yang and Land, 2006) to cleanly identify the effect of birth
cohorts.
3.3 Unobserved country time-invariant heterogeneity
In order to test Hypothesis 1 to 3 it is indispensable to control for unobserved
country time-invariant heterogeneity.9This especially applies to Hypothesis 1
and 2 where the variables of main interest are measured on a country-level.
Giventhatonenevercanbesurethatallrelevantcontrolvariablesareincluded,
an estimation without country fixed-effects is less convincing, since unobser-
vable factors may be correlated with the variables of main interest. The
importanceofcountryfixed-effectscan easily bedemonstrated. Ifweaugment
specifications (Ia) and (Ib) with country fixed-effects, neither the coefficient of
themeasureofthesizeofthewelfarestate(seecolumn(IIa)inTable1),norof
the unemployment rate (see column (IIb) in Table 1) exerts any statistical
significance. This suggests that both variables are correlated with unobserved
factors.10Therefore, we will control for country fixed-effects in all our
specifications below.
3.4 Socio-economic control variables
FH controls for the following socio-economic characteristics: age, sex, family
status, labormarketstatusand household income.Wethink thatthisbasicset
of socio-economic control variables should be augmented by any measure of
educational attainment.11The level of educational attainment is correlated
with the level of public social expenditures, the unemployment rate and the
birth cohort. Not controlling for educational attainment could result in
omitted variable bias and misleading conclusions with respect to Hypothesis
1 to 3. Moreover, we suggest a more detailed specification of labor market
status and split up the status non-employed in ‘unemployed’ and ‘out of labor
9In principal, one would prefer to control for unobserved individual time-invariant heterogeneity.
However, to our best knowledge there is no panel data on benefit morale available.
10FH includes five country-level control variables (ethical fractionalization, latitude, legal origin, tax
decentralization and autonomous regions) which effectively group together countries and mitigates the
omission of country fixed-effects to some degree.
11To capture the respondents’level ofeducational attainment we use the school leavingage. For detailed
information please refer to the Data Appendix in Halla, Lackner and Schneider (2009).
r 2010 Blackwell Publishing Ltd.
63
THE CASE OF BENEFIT MORALE

Page 10

force’ and compare them to the base group of employed individuals. Finally,
we suggest including two additional control variables: the number of children
and the size of the place of residence.
This augmented specification of socio-economic control variables provides
additional insights, see specifications (IIIa) and (IIIb) in Table 1. The level of
education exerts a positive and the size of the place of residence a negative
statistically significant effect on benefit morale. Further, it becomes clear that
the negative impact of non-employment is mainly due to unemployment, and
only to a lesser extent due to being out of the labor force.
3.5 Attitudinal control variables
Finally, there are a number of attitudinal variables for which one could easily
putforwardahypothesisontheirinterrelationtobenefitmorale.Forinstance,
FH controls for measures of religiosity, patriotism and confidence in parlia-
ment.Socialpsychologists(e.g.Tyler,2000)reasonablyarguethatpatriotism,
forexample,playsanimportantroleinshapingdeferencetoauthorities.People
who feel pride in society and in its authorities are more likely to obey those
authoritiesandtoaccepttheirdecisions.Accordingly,weexpectmorepatriotic
citizens to exhibit a higher benefit morale.
Is it therefore advisable to control for individual national pride when
testing Hypotheses 1 to 3? In our view, it can be misleading to control for
attitudinal control variables. For instance, individual patriotism may also be
influenced by public social spending. Conditioning on it would tamper
with that part of the causal effect of a public social spending on benefit
morale that operates through patriotism. If we would add more attitudinal
controlvariables(whicharetypicallyhighlycorrelatedwithbenefitmorale)we
would eliminate further potential channels through which public social
spending affects benefit morale. Moreover, we would not gain any further
insights if we find evidence that certain attitudinal control variables are
statistically significant determinants of benefit morale. Firstly, these coeffi-
cientscouldonlybeinterpretedascorrelations.Ineachcase,onehastosuspect
asimultaneitybias.Forinstance,wewouldnotonlyexpectthatmorepatriotic
citizens have a higher level of benefit morale, but also vice versa, meaning
that citizens with a high benefit morale exhibit higher levels of national
pride. Finally, these attitudinal variables are not easily amenable to policy
interventions.
We find that the exclusion of measures of religiosity, patriotism and
confidence in parliament does not change the qualitative results (detailed
output available upon request) and abstract from attitudinalcontrol variables
in the following empirical analysis.
64
r 2010 Blackwell Publishing Ltd.
MARTIN HALLA/MARIO LACKNER/FRIEDRICH G. SCHNEIDER

Page 11

4 ESTIMATION STRATEGY
Inordertoaccountforthemethodologicalissuesdiscussedaboveweestimatea
cross-classified fixed-effects model (Yang and Land, 2006),
BMi;j;c;t¼a0þ
X
12
2
l¼0
ZlSct?5lþ
X
periodtþxX þ pY þ
2
k¼0
zkUct?5kþ b1agei;j;c;tþ b2age2
i;j;c;t
þ
X
X
j¼2
18
cohortjþ
X
4
t¼2
X
10
m¼2
ymincomem;i;j;c;t
þ
c¼2
countrycþ ei;j;c;t;
ð1Þ
where BMi,j,c,tstands for the benefit morale of individual i, of birth cohort j,
fromcountrycinperiodt.Theinformationonbenefitmoraleisbasedonwave
two to four of the E/WVS. After cleaning the data we have information on
more than 30,500 respondents from 18 OECD-member countries for several
years between 1990 and 2000.12
InordertotestHypothesis1weincludepublicsocialexpendituresmeasured
asapercentageofGDPSct-5linthecurrentperiod(l50),withalagoffiveyears
(l51) and with a lag of ten years (l52). To test Hypothesis 2 we equivalently
examine the effect of past and current unemployment rates Uct25l. In order to
disentangletheeffectsofage,periodandbirthcohort(Hypothesis3)weinclude
age, age2, a series of binary variables cohortjcapturing 13 different groups of
birth cohorts as suggest by (Yang and Land, 2006), and period fixed-effects
periodt.
Inaddition, we include ona country level macroeconomic controlvariables
X (GDP per capita and GDP-deflator from OECD sources) and on an
individual level a set of socio-economic control variables Y included in the
E/WVS, comprising the respondent’s sex, marital status (married or not),
numberofchildren,sizeofplaceofresidence(measuredonathree-pointscale),
education(capturedbytheschoolleavingage),labormarketstatus(employed,
self-employed, unemployed, out of labor force) and the household income
(measured on a ten-point scale).13With respect to household income we
12Among otherswe have delete4,712 observations withstrictly impossible combinations of respondents’
age and year of birth. The number of available observations per country and year, and the respective
average levels of benefit morale can be found in Table 2 and Table 3 in Halla, Lackner and Schneider
(2009).
13Some respondents have reported an unrealistically low or high school leaving age. We have decided to
restrict the school leaving age to be within the age of 10 and 28 and disregard 4,263 observations. For
details please refer to the Data Appendix in Halla, Lackner and Schneider (2009).
r 2010 Blackwell Publishing Ltd.
65
THE CASE OF BENEFIT MORALE

Page 12

suggest a very flexible specification and include a binary variable incomem,i,j,c,t
for each income category m, where the base group is equal to the group with
lowest household income. This specification does not impose any functional
form on the effect of household income on benefit morale. Finally, we control
for country fixed-effects countrycto allow for unobserved country-specific
Table2
Determinants of benefit morale: testing hypotheses 1 to 3.a
(I)(II)(III)(IV)
HYPOTHESIS 1
Public social expend.t
0.0297???
(0.0055)
0.0205
(0.0125)
2 0.2212???
(0.0267)
0.0313???
(0.0099)
0.0429???
(0.0025)
2 0.0613???
(0.0146)
2 0.3646???
(0.0089)
0.0354???
(0.0112)
Public social expend.t25
Public social expend.t210
2 0.1204???
(0.0223)
2 0.1775???
(0.0427)
HYPOTHESIS 2
Unemployment ratet
0.0625???
(0.0201)
0.0945???
(0.0156)
2 0.0887???
(0.0178)
0.0098
(0.0142)
0.2581???
(0.0194)
0.1338???
(0.0070)
2 0.2557???
(0.0119)
0.0680?
(0.0353)
Unemployment ratet25
Unemployment ratet210
2 0.0615???
(0.0136)
2 0.1298???
(0.0416)
HYPOTHESIS 3
Birth cohort effectsb
COUNTRY-LEVEL CONTROL VARIABLES
GDPtp.c. (in $100,000)
yesyesyes yes
2 0.0003
(0.0017)
2 0.0057???
(0.0017)
0.0303???
(0.0035)
2 0.0201???
(0.0050)
2 0.0266???
(0.0013)
0.3045???
(0.0235)
yes
yes
0.0034
(0.0055)
GDPt25p.c. (in $100,000)
GDPt210p.c. (in $100,000)
2 0.0159?
(0.0083)
0.1992???
(0.0258)
yes
yes
GDP deflator0.3104???
(0.0308)
yes
yes
0.1848???
(0.0267)
yes
yes
Period fixed-effectsc
Country fixed-effects
INDIVIDUAL-LEVEL CONTROL VARIABLES
Age 0.0324??
(0.0125)
2 0.0002
(0.0001)
0.1706???
(0.0235)
0.1813???
(0.0261)
0.0021
(0.0117)
0.0128??
(0.0059)
2 0.0743??
(0.0303)
0.0323??
(0.0125)
2 0.0002
(0.0001)
0.1696???
(0.0235)
0.1813???
(0.0260)
0.0022
(0.0117)
0.0122??
(0.0058)
2 0.0739??
(0.0302)
0.0322??
(0.0125)
2 0.0002
(0.0001)
0.1707???
(0.0235)
0.1829???
(0.0261)
0.0021
(0.0117)
0.0131??
(0.0059)
2 0.0760??
(0.0305)
0.0319??
(0.0125)
2 0.0002
(0.0001)
0.1698???
(0.0235)
0.1816???
(0.0260)
0.0024
(0.0117)
0.0127??
(0.0058)
2 0.0742??
(0.0303)
Age2
Female
Married
No. of children
School leaving age
Size of town
66
r 2010 Blackwell Publishing Ltd.
MARTIN HALLA/MARIO LACKNER/FRIEDRICH G. SCHNEIDER

Page 13

time-invariantheterogeneity.Allstandard errorsare clusteredbycountry and
year (Moulton, 1990).
Sincebenefitmoraleismeasuredonaten-pointscale,itisstrictlyspeakingan
ordinal measure, which requires an ordered response model. Nevertheless,
since the qualitative results of ordered probit estimations turned out to be
equivalentandthescaleisratherlarge,wewillfortheeaseofpresentationfocus
on least squares throughout the paper.14
5 ESTIMATION RESULTS
OurmainestimationresultsarepresentedinTable2.InordertotestHypotheses
1 and 2 we include the levels of public social spending and the unemployment
ratewithalagoftenyears,withalagoffiveyearsandinthecurrentperiod.This
specification allows for different short, medium and long run effects of both
variables on benefit morale. As column (I) shows, both a higher level of public
social spending and a higher unemployment rate in the current period have a
quantitatively small positive effect on benefit morale. An increase in public
socialspendingbyonepercentagepoint(samplemeanisequalto20.83percent)
Table2. (Contd)
(I)(II)(III)(IV)
Self-employedd
2 0.0318
(0.0518)
2 0.0167
(0.0429)
2 0.3897???
(0.0671)
yes
8.0434???
(0.5835)
2 0.0358
(0.0518)
2 0.0145
(0.0427)
2 0.3884???
(0.0672)
yes
9.4366???
(0.5299)
2 0.0325
(0.0519)
2 0.0182
(0.0427)
2 0.3927???
(0.0673)
yes
9.6978???
(0.7340)
2 0.0359
(0.0518)
2 0.0142
(0.0426)
2 0.3882???
(0.0674)
yes
10.0476???
(0.5713)
Out of labor forced
Unemployedd
Incomee
Constant
R-squared0.0960.0950.096 0.096
aThedependentvariableisequaltobenefitmoralemeasuredonaten-pointscale,wherehighervalues
indicate a higher level of benefit morale. Estimated using ordinary least squares. The number of
observationsisineachestimationequalto30,582.Robuststandarderrors(allowingforclusteringby
country-years) in parentheses below.
?,??and???indicatestatisticalsignificanceatthe10-percentlevel,5-percentlevel,and1-percentlevel,
respectively.
bCoefficients are displayed in Figure 1.
cThe base group is equal to 1990.
dThe base group is equal to employed individuals.
eCoefficients are displayed in Figure 2.
14AspointedoutbyAiandNorton(2003)theinterpretationofnonlinearmodelsisquitecumbersomeand
not fully demonstrative.
r 2010 Blackwell Publishing Ltd.
67
THE CASE OF BENEFIT MORALE

Page 14

of GDP is associated with an increase in benefit morale by about 0.03 points.
An equivalent rise in the unemployment rate (sample mean is equal to 7.52
percent) increases benefit morale by about 0.06 points. These positive effects
are contrary to what Hypotheses 1 and 2 predict. However, our estimates of
the lagged values show that after a certain period of time an adverse effect f
both variables kicks in. In the case of public social spending we observe no
mediumruneffect,butaquantitativelyimportantnegativelongruneffect.An
increase in public social spending by one percentage point today is estimated
to decrease benefit morale by about 0.22 points ten years later. Likewise, we
observe a detrimental long run effect of high unemployment on benefit
morale.Accordingtoourestimationanincreaseintheunemploymentrateby
one percentage point decreases the level of benefit morale ten years later by
about minus 0.09 points.
Ifweleaveoutthefiveyearlag ofbothvariables(seecolumn(II)in Table2)
the negative long run effect of a more pronounced welfare state remains
present, though the effect turns out to be a bit smaller in both cases. Notably,
duetotheexclusionofthefiveyearlagofbothvariablesthepositiveeffectofthe
currentunemploymentratevanishes.Theeffectsofallother(control)variables
arerobustduetothismodification.Totesttherobustnessofourresultswehave
augmentedourtwospecificationsbyGDPwithalagoffiveyearsandwithalag
of ten years as additional control variables, see columns (III) and (IV).
Comparing specification (II) with specification (IV) we see that this extension
has no impact on the qualitative results, however, it tends to increase the
estimates of the negative long run effects. Similarly, we observe an increase in
all effects in absolute terms in specification (III) compared to specification (I).
Moreover, the adverse effect of public social spending kicks in already in the
medium run.
OfcoursethelagoffiveandtenyearswithrespecttoHypotheses1and2are
no natural choice, however, it turned out to be the most robust specification
amongall.15Ideally,onewouldliketoestimatetheeffectoflaggedpublicsocial
spending and unemployment for each year over the preceding decade. This
would allow to trace out the full adjustment path of benefit morale. However,
duetothe smallnumberofcountry-years this can not beimplementedyetand
one has to wait for further waves of the E/WVS to be conducted.
Insum,allourdifferentspecificationssuggestthatbothanincreaseinpublic
social expenditures and a rise in unemployment have small positive short run
(or no) effects on benefit morale. These, however, are (partly) crowded out by
adverse medium and long run effects. In the case of public social expenditures
thenegativemediumandlongruneffectsclearlydominatethepositiveeffectin
15Notably, the results with respect to Hypothesis 3 (see below) are completely unaffected by these
modifications.
68
r 2010 Blackwell Publishing Ltd.
MARTIN HALLA/MARIO LACKNER/FRIEDRICH G. SCHNEIDER

Page 15

the current period. This result is in line with the idea that individuals have to
experiencegenerouswelfarearrangementsforquitesometimeuntiltheyadapt
theirsocialnormtowardsacceptingbenefitfraud,oratleastconsideringittobe
aminoroffence.Wetherefore,interpretthisasevidenceinfavorofHypothesis
1 and to a lesser extent of Hypothesis 2 and corroborate the theoretical
literature which suggests that disincentive effect of welfare arrangements
may materialize only with considerable time lags. Our results also affirm the
basic empirical finding of FH, in the sense that it boils down to a disincentive
effect of different welfare arrangements on benefit morale. However, our
analysis highlights that this relationship is characterized by a specific dynamic
structure which requires sensible econometric modeling.
Basedonourcross-classifiedfixed-effectsmodelwedonotfindanyevidence
for birth cohort effects (Hypothesis 3). The four charts in Figure 1 display the
estimated coefficients for the birth cohorts effects for the four different
specifications summarizedin Table2.Theyshowthatfor eachbinary variable
capturing a birth cohortgroup,the 95percent confidenceinterval includes the
value zero. This means, that none of the birth cohort group dummies is
statisticallysignificantdifferentfromzeroatconventionallevels.Thisfindingis
veryrobustandholdsacrossallspecifications.Anequivalentanalysiswasalso
carried out using several other cohort structures (i. e. other than the five-year
groups). We also substituted age and age2with a series of binary variables
capturingdifferentagegroups.Inanycase,wefindnosignificanteffectofbirth
cohorts on benefit morale.
Figure1
Cohort effects on benefit morale
r 2010 Blackwell Publishing Ltd.
69
THE CASE OF BENEFIT MORALE