Trustees of Princeton University

Trustees Prcet University Earngs Inequality Welfare Spendg: A Disaggregated Analysis Author(s): Karl Ove Moene Michael Wallerste Source: World Politics, Vol. 55, No. 4 (Jul., 2003), pp. 485-516 Published : Cambridge University Press Stable URL: http://www.jsr.org/stable/25054236. Accessed: 07/05/2011 09:05 Your use JSTOR archive dicates your acceptance JSTOR's Terms Cditis Use, available at. http://www.jsr.org/page/fo/about//terms.jsp. JSTOR's Terms Cditis Use provides, part, unless you have obtaed prior permissi, you may not download an entire issue a journal or multiple copies articles, you may use ctent JSTOR archive ly for your persal, n-commercial use. Please ctact publisher regardg any furr use this work. Publisher ctact formati may be obtaed at. http://www.jsr.org/acti/showpublisher?publishercode=cup.. Each copy any part a JSTOR transmissi must cta same copyright notice appears screen or prted page such transmissi. JSTOR is a not-for-prit service helps scholars, researchers, students discover, use, build up a wide range ctent a trusted digital archive. We use formati technology ols crease productivity facilitate new forms scholarship. For more formati about JSTOR, please ctact support@jsr.org. Cambridge University Press Trustees Prcet University are collaboratg with JSTOR digitize, preserve extend access World Politics. http://www.jsr.org

EARNINGS INEQUALITY AND WELFARE SPENDING A Disaggregated Analysis By KARL OVE MOENE MICHAEL WALLERSTEIN* I. Introducti collect spend average around 45 percent GOVERNMENTS GDP advanced dustrial societies, about half govern ment goes fund various transfer pay ments services cstitute what is commly called welfare state. Perhaps most comm view welfare is se are outcome a lg political struggle which workers ir allies used power ballot box obta some redress for equalities generated market. In words Huber Stephens: "The struggle welfare states is a struggle distributi, thus organizatial power those stg benefit from re distributi, workg lower middle classes, is crucial."1 Or scholars have emphasized political fluence beneficiaries welfare who are outside labor market, such as elderly.2 But wher key groups are defed class positi, come, or age, most scholars have viewed welfare redistributive terms. The redistributive view welfare as policy, formalized a series papers Romer, Roberts, Meltzer Richard, implies higher market comes generates higher levels political * We thank Nolan McCarty three anymous referees for comments. helpful We thank MacArthur Foundati Norwegian Research Council for fancial support. 1 Evelyne Huber John D. Stephens, Political Choice Global Markets: Development Crisis Advanced Welfare States (Chicago: University Chicago Press, 2001), 17. For or recent books emphasize centrality eir political mobilizati or ecomic organizati workers for explag cross-natial differences size welfare state, see Alexer Hicks, Social Democracy Welfare Capitalism: A Century Income Security Politics (Ithaca, N.Y.: Cornell University Press, 1999); Harold Wilensky, Rich Democracies (Berkeley: University California Press, 2002); Duane Swank, Global Capital, Political Institutis, Policy Change Developed Welfare States (Cam bridge: Cambridge University Press, 2002). 2 Fred C. Pampel John B. Williams, cies 1950-1980," American Journal Sociology "Welfare Spendg 93 (May 1988). Advanced Industrial Democra World Politics 55 (July 2003), 485-516

486 WORLD POLITICS support for redistributive.3 The basic tuiti is low come earners have more ga less lose than do perss with high comes from expansis welfare. Thus, poorer majority voters relative average come, greater ex pected support for welfare In. e-dimensial model over votg welfare where voter with median come is decisive, statistic is ratio median come mean key come. The more skewed distributi come or, more precisely, lower ratio median mean come, higher level welfare desired a majority voters. Welfare pol icy is expected "lean agast wd" sense greater pre-tax transfer come, greater elecral sup port for government redistribute from rich poor. An alternative view welfare state is social-surance provide surance rar than redistributi. As Baldw observed a study origs welfare state five European countries, "Pro tecti agast risk has been more sought universally than a redistribu ti resources."4 Of course, all surance are redistributive sense fire surance redistributes resources from those lucky never a enough experience fire ir house those who have misfortune experiencg such. Neverless, fire surance is not re distributive ex ante. We do not expect fire surance be more popular amg poor than amg rich.5 The typical social-surance program, however, is neir pure redis tributi nor pure protecti agast risk but rar a mixture 3 Thomas Romer, "Individual Welfare, Majority Votg, Properties a Lear Income Tax," Journal Public Ecomics 14 (February 1975); Kev W. S. Roberts, over "Votg Income Tax Sched ules," Journal Public Ecomics 8 (December 1977); Allan H. Meltzer Scott F. Richard, "A Ra tial Theory Size Government,"Journal Political Ecomy 89 (Ocber 1981). 4 Peter Baldw, The Politics Social Solidarity: The Class Bases European Welfare State (Cam bridge: Cambridge University Press, 1990), 18. For a survey ecomic arguments favor public provisi surance, see Nicholas Barr, "Ecomic Theory Welfare State: A Survey Interpretati," Journal Ecomic Literature 30 (June 1992). Iversen Cusack terpret welfare state as surance agast risk come loss occasied shift jobs from manu facturg service secr, while Rodrik Garrett terpret welfare state as surance agast risks entailed creased ternatial ecomic see tegrati; Torben Iversen Thomas R. Cusack, "The Causes Welfare State Expansi: Dedustrializati or Globalizati?" World Poli tics 52 (April 2000); Dani Rodrik, "Why Do More Open Ecomies Have Larger Governments?" Journal Political Ecomy 106 (Ocber 1998); Gefrey Garrett, Partisan Politics Global Ecomy (Cambridge: Cambridge University Press, 1998). The redistributive surance framework used this paper was first presented Karl O. Moene Michael Wallerste "Inequality, Social In surance Redistributi," American Political Science Review 95 (December 2001). 5 The strg correlati exists between social-surance as a share GDP GDP per capita data sets clude both high-come low-come countries suggests richer voters a prefer spend larger share come social surance; correlati is documented Harold Wilensky, The Welfare State Equality (Berkeley: University California Press, 1975). An alternative deadweight explanati is capacity governments collect revenues without costs rises with ecomic development. imposg large

EARNINGS INEQUALITY & WELFARE SPENDING 487 two. Social-surance advanced dustrial societies generally provide surance comm agast risks terms are more favor able for low-come dividuals than for high-come dividuals. In this article, we suggest neir redistributi nor surance ale can expla how come affects dem for social sur ance. We argue, stead, it is mixture two motives?a mixture differs from e social-surance policy anor? determes relatiship between distributi come support for welfare. We beg showg extendg framework Romer Meltzer Richard clude provisi surance redistributive terms generates predictis ccerng impact equality support for social-surance depend mixture redistributi surance, design. We n cduct an empirical examati impact earngs equality welfare disaggregated pensis, health care, surance agast unanticipated come loss, family bene fits, housg subsidies, poverty alleviati. We fd little or no rela tiship between earngs as a share GDP for pensis, health care, family benefits, means-tested poli cies. In ctrast, we fd significantly lower countries with higher earngs for welfare surance provide for workers who have lost ir come because layfs, ill health, or accidents, cstitute about 30 percent tal social surance. Instead "leang agast wd," a substantial share welfare is better characterized as "bendg wd," is, as declg creases. A simple model vot g over redistributive surance predicts exactly this pattern. Our study is not first present evidence relatiship between come social-surance ad vanced dustrial societies is not csistent with a purely redistributive model. at OECD Lookg countries, Rodriguez Moene Waller ste found higher come be associated with lower social surance as a share GDP.6 In United States, Rodriguez found no relatiship between welfare at state level, while Mfitt, Ribar, Wilhelm found Aid for Dependent Children (AFDC) be lower states where distri buti come was most a unequal.7 Usg broader sample fifty 6 Francisco Rafael Rodr?guez Caballero, "Essays Political Ecomy Inequality, Redistri buti Growth" (Ph.D. diss., Harvard University, 1998); Moene Wallerste (fn. 4). 7 Rodriguez (fn. 6); Robert Mfitt, David Ribar, Mark Wilhelm, "The Decle Welfare Benefits U.S.: The Role Wage Inequality,"Journal Public Ecomics 68 (June 1998).

488 WORLD POLITICS rich poor countries, Perotti found no significant relatiship be tween social-surance.8 In ctrast se studies, Milanovic fds more unequal countries redistribute more a sample cludes advanced dustrial societies newly dustrializg countries East Asia, where redistributi is measured difference between pre-tax transfer Gi coefficient post-tax transfer Gi coefficient.9 Sce social-surance pro grams are not ly redistribute come, studies re distributi general studies social-surance particular may arrive at different cclusis. In this article we study relatiship between earngs major compents social-surance. Our c tributi is highlight differences amg ma categories social surance relatiship between come ex penditures, differences are obscured studies aggregate welfare let, ale studies redistributi general. In next secti we describe a simple model over votg redistributive sur ance different comparative static results ccerng come equality support for for different categories social-surance. In Secti III we analyze disaggregated data social-surance OECD countries show relatiship between social-surance varies amg social-surance a manner matches model Secti II. In fal secti we summarize our fdgs discuss or possible differences we observe explanatis amg social surance programs relatiship between expen ditures. The pros claims made Secti II are presented Appendix 1. II. Theoretical Framework The argument this article rests a simple model voters politics: are assumed have well-defed, strictly sgle-peaked preferences over level fundg for each various social-surance poli cies depend each voter's come relative average come. With strictly sgle-peaked preferences depend voters' come, 8 Rober Perotti, "Growth, Income Distributi Democracy: What Data Say," Journal Ecomic Growth 1, no. 1 (1996). 9 Branko Milanovic, "The Median-Voter Hyposis, Income Inequality, Income Redistribu ti: An Empirical Test with Required Data," European Journal Political Ecomy 16, no. 2 (2000).

EARNINGS INEQUALITY & WELFARE SPENDING 489 voters with median come are pivotal sense ir preference between any two alternatives is always shared a majority elecrate. In such an envirment, it is natural identify ideal policy median voters as political equilibrium. Elecral competiti between two or two parties blocs as parties, Scavian-style multiparty systems where parties organize socialist bourgeois blocs, forces both right left compete for support voters around median come distributi. Regardless which party ws electi, policy is adopted is close policy preferred median group voters. This highly stylized model politics social surance can be criticized many grounds. Voters, it is claimed, are generally ill formed about policy choices.10 Instead, many have argued, welfare policy is determed political fluence labor movement,11 policy novatis bureaucrats,12 or even preferences em In ployers.13 formal literature two-party, e-dimensial model redistributive politics has been extended clude more than two parties14 more than e policy dimensi.15 Neverless, this article we rely e-dimensial, two party model redistributive politics developed Romer Meltzer Richard, extended cover social surance. While unis, bureau crats, employers have all played important roles negotiatg details social-surance, end social-surance are or not or adopted parties coalitis parties manage w a majority votes. Voters may know little or nothg about details policy choices facg legislars, but if voters vote retrospectively, rewardg cumbent government if ir welfare has creased punishg cumbent orwise, parties government have a strg elecral centive adopt raise welfare a 10 Dald Kder, "Diversity Complexity American Public Opi," Ada Fifter, ed., Political Science: The State Disciple (Washgt, D.C.: APSA, 1983). 11 Walter Korpi, The Workg Class Welfare Capitalism (Ld: Routledge Kegan Paul, 1978); Hicks (fn. 1); Huber Stephens (fn. 1); Wilensky (fn. 1). 12 Hugh Heclo, Modern Social Policies Brita Sweden (New Haven: Yale University Press, 1974). 13 Peter Swens, Capitalists agast Markets (Oxford: Oxford University Press, 2002). 14 David Austen-Smith, "Redistributg Income under Proportial Representati," Journal Po litical Ecomy 108 (December 2000). 15 John Roemer, "Why Poor Do Not Expropriate Rich," Journal Public Ecomics 70 (De cember 1998); idem, Political Competiti (Cambridge: Harvard University Press, 2001); David Austen-Smith Michael Wallerste, "Redistributi a Divided CMS-EMS Society" Discussi Paper 1362 (Evanst, 111.: Northwestern University, 2003).

490 WORLD POLITICS majority voters.16 For se reass, we believe policy pre ferred a majority voters all alternatives, when such a ex policy ists, is an important (albeit not necessarily ly) determant are adopted. But before framework developed Romer Meltzer Richard can be expected expla poli tics social surance, model must be modified clude risk. Csider an elecrate composed self-terested, risk-averse vot ers who differ ir come when employed but face a comm risk losg ir employment next In we period. particular, will rely followg assumptis: 1. The wage distributi is lognormal. Let a2 denote variance log wages. Sce we will csider changes ( is, a2), holdg average wage cstant, re is no loss generality assumg average wage e. equals 2. All voters receive a known wage with probability n. There is, however, a nzero probability, (1? 7r), each voter will lose his or her come because unemployment, jury, or illness. To keep model as simple as possible, probability beg employed, 7U, is assumed be same for all voters.17 3. Voters are assumed be identical terms ir aversi risk. As our measure voters' a willgness accept lower average come for exchange less uncerta come, we use coefficient relative risk aversi, a = -cu"(c)/u'(c) where u{c) represents voters' over c. The preferences csumpti, higher (i is, more voters are willg pay for surance agast loss a fracti ir come. We assume given ui is same for all voters \x > 1, which implies dem for surance rises as come creases.18 4. Social-surance /, can take are fanced a flat tax wages, denoted any value between zero some Taxati is assumed?max.19 16 Chrispher Achen Larry Bartels, "Ignorance Bliss Democratic Politics: Party Com petiti with Unformed Voters" at (Paper presented annual meetgs Midwest Political Science Associati, Chicago, April 25-28,2002). 17 We discuss relaxg this assumpti below. 18 The assumpti u is cstant is made simplify mamatical expressis, but it is not necessary. The assumpti [i > 1 is critical for our results. Both u are assumptis regardg sup ported studies allocati household savgs. See Irw Friend Marshall E. Blume, "The Dem for Risky Assets," American Ecomic Review 65 (December 1975). 19 We have made modelg choice represent differences social-surance terms differences distributi benefits, rar than terms differences tax fances benefits. A more general approach would be defe post-tax transfer come as a functi pre?tax transfer come, as John Roemer, "Does Democracy Engender Equality," Edward Mansfield Richard Siss, eds., Political Knowledge Public Interest (Columbus: University Ohio Press, 2003). For purposes this paper, however, assumpti a flat tax is a reasable approximati. In most countries we study, much welfare budget is fanced a payroll tax is usually flat. (Denmark is an outlier relyg almost exclusively come value-added taxes.) Moreover, a recent study progressiveness persal come tax twelve OECD countries WagstafFet al., found "no lk between pre-tax degree redistributi brought about persal come tax"; see Adam et Wagstaff al., "Redistributive Effect, Progres sivity Differential Tax Treatment: Persal Income Taxes Twelve OECD Countries," Journal

EARNINGS INEQUALITY & WELFARE SPENDING 491 impose a deadweight cost which we model implicitly assumg tal tax revenues per capita, Ty are given a twice differentiable functi tax rate, x(/), multiplied average earngs, n (sce fracti n are workg average wage is e), or = T(t) nx(t). The functi x(/) is assumed satisfy followg properties: (1) x(0) = 0 (no taxes are collected when tax rate is zero); (2) x'(0) = 1 (re is no deadweight loss at / = 0); (3) x"(t) < 0 ( dead weight cost taxati rises as tax rate creases); = (4) x'(?max) 0 for some / < 1 (re is some tax rate / < max v max 1 beyd J which furr creases tax rate do not crease tax revenues). 5. Social-surance are two represented functis, ^ 0 be(w,t)? bn(w,t) 0, where be(w,t) represents transfer payment received an em worker who earns a ployed wage w bn(w,i) represents transfer pay ment received a worker without when employment tax rate is /. Note benefit may be a functi worker's wage or, case a worker without current employment, worker's past wage. Voters' preferences regardg rived from ir expected utility social-surance are de Eu = nu(ce) + (1 - n)u(cn) (1) where c? is voter s post-tax transfer come when employed cn is voter s come when not employed. Posttax, posttransfer come when employed is equal posttax wage (1 - t)w plus welfare benefit, or c = (1 - t)w + Voters who are not be(w,t). workg receive cj.w) = bn(wyt). A voters ideal policy is tax rate or (1) subject budget cstrat I [nbe(w,t) + (1 - n)bn(w,t)]f(w)dw level maximizes = nx(t) (2) where f(w) is probability density functi wage distributi. Equati 2 states average benefits received those em ployed those not employed at each wage level must equal tax rev enues per capita. Different social-surance can be represented different specificatis functis be{w,t) bn(w,t). Sce voters have stricdy sgle-peaked preferences with all specificatis be(wyt) we bn(wyt) exame, tax rate or benefit level preferred voter who receives median wage is a preferred majority Public Ecomies 72 (April 1999), 83. Thus, neglectg cross-natial differences redistrib utive impact come tax, we do not appear be a neglectg facr is systematically related pretax come.

492 WORLD POLITICS any or alternative. Therefore, we identify preferred policy voter with median come as political equilibrium. Csider, as a benchmark, case simple which all benefits are as paid equal payments employed workers, or = be(w,t) b(t) bn{wyi) = 0. The budget cstrat implies b(t) = x(/). This is a purely redistributive policy no surance provides agast job loss. In this case, an crease wage creases equilibrium level as, stated first claim:?claim 1. When be(w,t) = b(j) = bn{w,t) 0, equilibrium tax rate benefit level rises as wage distributi creases. The pro this all or claims are presented Appendix 1. This claim simply reproduces ma result Romer Meltzer Richard model votg over redistributive. An crease variance a lognormal distributi, holdg mean cstant, a implies decle median come. From pot view median voter, a given benefit can now be obtaed at a lower price sce median voter's share tax burden decles as median voter's come falls. Hence, median voter a prefers higher level. Social-surance, however, do not pay benefits currently employed workers ly. Social-surance eir target benefits those who are not currently employed (such as unemployment surance) or provide benefits everye (such as health care). Csider, first, family provides come replacement for those who have lost ir earngs due unforeseen circumstances such as layfs, sickness, or accidents. The benefits from come-replacement are received ly workers without current employment, im plyg be(w,t) = 0. In additi, benefits are typically tied past earn a gs accordg redistributive formula we write as = bn(wyt) [^ + (1 -?)w] b(t) where 0 < \ ^ l.20 In or words, come-replacement 20 The assumpti are come-replacement redistributive, or cj > 0, is csistent with way most social-surance programs are designed. For example, average replacement ratio for unemployment surance, notati bn{w,t)/w paper, is 18 percent higher for a worker who receives two-thirds median wage than for a worker who receives median wage countries our data set; Organizati for Ecomic Cooperati Development, Dataset Bene fit Replacement Rates (Paris: OECD, no date). Of course, it would be preferable if > \ 0 was a cclusi rar than an assumpti. However, e-dimensial model no politics lger applies when both ^ b are chosen simultaneously. For a model where? is chosen at a "cstitutial" stage maximize a social welfare functi while b(t) is chosen self-terested voters a secd "elecral" stage, see Georges Casamatta, Helmuth Cremer, Pierre Pestieau, "Political Sustaability Design Social Insurance," Journal Public Ecomics 75 (March 2000).

EARNINGS INEQUALITY & WELFARE SPENDING 493 are assumed provide an come floor h>b(t) plus frac - ti (1 Qb(t) past earngs. The term b(t) is average benefit re ceived those who receive benefits. Sce taxes are collected from fracti n populati while benefits are paid fracti (1-7i), budget cstrat implies : b(t) = [n /(l - 7t)]t(/). In ctrast case where benefit is paid those who are em an ployed, crease reduces dem for welfare spend g when benefit is received those without as employment, stated secd claim.?claim 2. When be{w,t) = 0 = bn{w,t)? + - (1 Qw\b{t) where 0 <? <s 1, equilibrium tax rate benefit level decles as wage distributi creases. In this case a reducti come median voter has two effects work opposite directis. As previous case, price a given level benefits for median voter decles, re creasg median voter s dem for. But now re is an come effect pushes opposite directi. A decle come median voter reduces amount surance median voter wishes purchase. We demstrate Appendix 1 come effect domates price (or substituti) effect this case, which implies support for benefits ment decles as creases. for those without employ Neir pure surance nor pure redistributi can expla why risg earngs lowers political support for come-replacement amg voters with below-average earngs. The key is mix ture surance redistributi. When come-replacement poli cies are redistributive ( is, when? > 0), a change an dividuals earngs, holdg average earngs cstant, duces a less than pro portial change social-surance benefit would be received event unemployment, illness, or accident. Workers whose earngs have risen relative social-surance benefit prefer crease benefit a little, even at cost a tax. higher Cversely, workers whose earngs have fallen relative social-surance benefit prefer reduce benefit a litde order resre some ir posttax come when workg. Thus, a mean-preservg decle raises wage majority workers who earn less than average also raises level social-surance majority prefers. In ctrast social surance agast unforeseen come loss, social surance programs like health care are universalistic sense

494 WORLD POLITICS benefit is same at all come levels is received regardless wher or not beneficiary is currently employed. A reasable char acterizati programs like health care is = = simply bn(w,t) be(w,t) b(t).21 Pensis are like health care unlike unemployment surance sense public pensi systems provide come up reachg retirement age all workers. Unlike health care, however, pensis de pend up earngs, typically with a redistributive formula for calculat g benefits.22 Pensis, refore, might be represented bn{w,t) = be(wyt) = [^ + (1 - t)w] b(t)y with 0 < 2; < 1. The third claim covers both programs such as health care, where?, = 1, pensis, where? < 1.?Claim 3. When be(w,t) = = bjw,t) [? + - (1 Qw]b(t) where 0 <??s 1, equilibrium tax rate benefit level may eir rise or decle as wage distributi creases. In particular, equilibrium tax benefit crease as creases if coefficient relative risk aversi, i, is sufficiently close e, while equilibrium tax benefit decles as equality creases if i is sufficiently large. Wher benefits are same for all = (^ 1) or depend earngs (? < 1), equilibrium level benefits is an creasg functi when \x?> 1 a decreasg functi when i?? oo. For moderate levels risk aversi, effect can go eir way. The median voters preference for greater redistributi less surance as median come falls relative mean roughly balance each or. Fally, re are explicitly target poverty alleviati. Means-tested, which cstitute a mor part welfare budget but an important part budgets very poor households advanced dustrial societies, cannot be examed a model self terested votg. The probability receivg payments targeted for poverty alleviati are zero virtually for a majority voters. Support for such must be based facrs such as altruism or fear crimal acts desperately poor. To summarize results can be derived from an extensi Romer Meltzer Richard framework clude risk job loss, relatiship between pretax come equilib rium level social-surance depends s policy design. A compressi wage distributi creases 21 The U.S. is exceptial devotg roughly 25 percent public health a family means-tested programs known as Medicaid; United States Bureau Census, Statistical Abstract United States 2001 (Washgt, D.C.: U.S. Government Prtg Office, 2001). 22 Walter Korpi Joachim Palme, "The Paradox Redistributi Strategies Equality: Welfare State Institutis, Inequality Poverty Western Countries," American Sociological Re view 63 (Ocber 1998).

EARNINGS INEQUALITY & WELFARE SPENDING 495 come workers with below-average earngs relative mean has two effects when surance is provided redistributive terms. The dem for surance rises while dem for redistribu ti falls amg majority workers whose come has creased. In case come-replacement, greater dem for surance domates reduced dem for redistributi support for social-surance creases. In case provide benefits for employed as well as for those without employ ment, enhanced dem for surance reduced dem for redistributi roughly cancel each or out. Therefore, redistributive social-surance targeted those who have lost ir come unexpectedly because layfs, sickness, or accidents is predicted be higher countries with more egalitarian distributis pre-tax transfer come. By ctrast, social surance benefits are received all workers is not predicted have a strg relatiship with e way or or. Our assumpti risk job loss is uncorrelated with earn gs is not critical. Ne results is altered allowg risk job loss depend a worker's positi come distributi provided worker with median ideal pot with regard social surance has below-average come. It is important note redistributive-surance framework does not imply high-wage workers desire higher come-replacement than do low-wage workers. The dem for surance depends risk as well as come. Low-wage workers may express greater support for unemployment surance than high-wage workers, for ex ample, sce probability beg laid f is higher for low-wage workers. What redistributive-surance framework implies is a workers dem for unemployment surance would crease if ratio workers come average come creased, c holdg stant risk job loss. In a comparis two countries with same distributi risk come loss but different distributis come, more skewed distributi come, lower level surance agast come loss desired a majority voters. III. Empirical Analysis In this secti we describe empirical relatiship between ma categories social-surance earngs eigh teen OECD countries from 1980 1995. We show relatiship between levels earngs varies across social- surance programs way is predicted extended frame

496 WORLD POLITICS work Romer Meltzer Richards. We beg with a discussi data used statistical analysis methodological is sues we cfrted. We n our present empirical results. Details regardg data sources can be found 2. Appendix Descripti Data OECD Accordg statistics, social-surance averaged 23 percent GDP 51 percent tal government ad vanced dustrial societies between 1980 1995.23 The welfare bud get can be divided three large categories three smaller categories. Pensis (old-age cash benefits) make up 30 percent welfare budget average. Public health csumes an av erage 26 percent welfare. Policies provide come support a wide variety circumstances (unemployment, disability, sickness, occupatial jury, death a cstitute 31 spouse) percent social-surance average. The remag 13 percent welfare budget is spent benefits services for families with children (9 percent welfare ), benefits targeted low come dividuals, refugees, digenous groups (3 percent wel fare ), housg subsidies (1 percent welfare ). Note smallness share explicitly dedicated poverty alleviati. Government for what is known as "welfare* U.S., is, programs which eli gibility for benefits is based primarily low come, averages ly 0.6 percent GDP advanced dustrial societies. Table 1 sum presents mary statistics, while Table 6 2 Appendix presents country means for each ma categories social-surance. In e-dimensial model over votg welfare, support for wel fare ratio come median voter depends mean come. Unfortunately, re are ly limited data ratio median mean come. However, OECD has pub lished data ratio at earngs different percentiles earngs distributi most OECD coverg countries from 1980 through 1995.24 The data refer annual come from wages salaries received full-time employees, both men women. We can use fact distributi wages salaries is well approximated log normal distributi write ratio median mean as 23 Organizati for Ecomic Cooperati 1980-1996 (CD-ROM) (Paris: OECD, 1999). 24 Organizati for Ecomic Cooperati idem, Employment Outlook (July 1996). Development, Social Expenditure Database Development, Outlook Employment (July 1993);

EARNINGS INEQUALITY 6c WELFARE SPENDING 497 median mean = exp(-av2) (3) where a2 is variance log wages salaries.25 The variance log wages, turn, can be derived from ratio wage at any two percentiles earngs distributi accordg formula a = k (w./w) (4) where w. is wage or salary received a worker at ith. percentile earngs distributi w. is wage or salary received a worker at /th percentile earngs distributi < withj /, k.. is a positive cstant depends i j. Equatis (3) (4) imply \n{w./w) is a reasable proxy for ratio median come mean.26 The OECD provides data 90/10,90/50, 50/10 wage ratios. As equatis 3 4 dicate, statistical results should not depend wage ratio is used. In practice, lognormal distributi is a good approximati but not a perfect characterizati actual distributi wages all variables are measured with error. There fore, we used all three available wage ratios our analysis. To save space, we ly report results usg 90/10 wage ratio, but our fdgs do not differ significantly when 90/10 wage ratio is re placed eir 90/50 or 50/10 wage ratio. Because wage salary data are not available an an nual basis for many countries because we do not thk small annual changes distributi come have an immediate political we impact, used average value 90/10 wage ratio for each five year period. That is, expla social-surance, say, 1985, we use average all measures 90/10 wage ratio are available for time period 1980-84. Thus, our data set csists data various social-surance programs as a share GDP eighteen countries years 1985,1990, 1995, with mea sures wage ( most or ctrol variables) averaged 25 For a discussi properties lognormal distributi its use as an approximati distributi come, see J. Aitchis J. A. C. Brown, The Lognormal Distributi (Cam bridge: Cambridge University Press, 1957). 26 Both our model data we use measure come refer wage earn salary ers who are eir workg full time or are temporarily without employment. Of course, not all voters fit se categories. Some work part time. Ors are permanently outside dependent labor force. To take all categories attachment labor market account would greatly complicate analysis, both oretically empirically.

498 WORLD POLITICS over time periods 1980-84,1985-89, 1990-94. We have fifty observatis after subtractg four cases which re is no mea sure wage with five-year time period.27 On average, a worker at 90th percentile received three times earngs a worker at 10th percentile. The most earn egalitarian gs distributi data set is Norway 1990-94, where ratio at earngs 90th at percentile earngs 10th was percentile less than two e. The least egalitarian earngs distributi was achieved U.S. 1985-89, when workers at 90th percentile received a wage or salary was 5.5 times earngs received workers at 10th percentile. As ctrol variables, we clude dependent variable e lagged period (five years), rate unemployment, share elderly voter populati, turnout, a measure cservative party partici pati government. We discuss each briefly turn. LAGGED DEPENDENT VARIABLE Budgetg is cremental. The best sgle predicr next periods welfare budget is current welfare budget. Indeed, simple regressi current tal social-surance past tal social-surance a (plus cstant) an yields R2 87.7 percent.28 Therefore, we clude lagged dependent variable set regressors. UNEMPLOYMENT RATE Once parameters unemployment surance are fixed, expendi tures unemployment benefits vary directly with rate unem ployment. Expenditures active labor-market even disability surance may also be sensitive rate. unemployment Thus, we clude rate unemployment same year as data when analyzg categories might be sensitive unemployment rate.29 SHARE OF ELDERLY IN THE POPULATION Government pensis health care may be affected share elderly populati, both because larger share elderly, greater need for mata elderly 27 The countries years data set are listed Appendix 2. 28 The regressi equati is yt = 3.03 +.938yM with R2 = 87.7 n = 50, where yt is tal welfare as a share GDP period stard error coefficient lagged dependent variable is.050. 29 The possible endogeneity unemployment is discussed below. /,

EARNINGS INEQUALITY & WELFARE SPENDING 499 reasable comfort because larger share elderly, larger share elecrate with a keen terest pensis. We use average share elderly populati previous five years (as is appropriate if share elderly primarily measures political strength elderly) rar than same year (as would be appropriate if share measures elderly primarily need) because five-year average fits data better than same year figure, although difference fit is small. TURNOUT Sce elecrate is not a representative sample adult popula ti as a whole, level turnout may affect support for welfare ex as penditures, argued Lijphart Franzese.30 Therefore, we clude average turnout electis lower house parliament (ex cept U.S., where we use turnout presidential electis) each five-year period. PARTISAN COMPOSITION OF GOVERNMENT The simple spatial model elecral competiti between two parties a competg sgle policy dimensi predicts two parties fer identical equilibrium absence uncertaty. In presence uncertaty about precise elecral csequences e ferg policy rar than anor, however, parties care about policy outcomes ( not just about wng) would propose divergent equilibrium.31 With uncertaty, positis parties care about policy choices a represent compromise between platform maximizes probability wng ( is policy preferred median voter) platform party would most like im plement. Therefore, we clude party power as a ctrol. Like many ors, we fd greatest partisan difference with respect welfare is which separates cservative parties from both center left we use parties.32 Accordgly, average share 30 Arend Lijphart, "Unequal Participati: Democracy's Unresolved Dilemma," American Political Science Review 91 (May 1997); Robert J. Franzese, Macroecomic Policies Developed Democracies (Cambridge: Cambridge University Press, 2002). 31 Roemer(fn. 15,2001). 32 Francis G. Castles, "The Impact Parties Public Expenditure" Castles, ed., The Impact Parties: Politics Policies Democratic Capitalist States (Ld: Sage Publicatis, 1982); G?sta Espg-Andersen, The Three Worlds Welfare Capitalism (Prcet: Prcet University Press, 1990). The tripartite divisi parties left, center, right follows Francis Castles Peter Mair, "Left-Right Political Scales: Some 'Expert' Judgements," European Journal Political Research 12 (March 1984). Socialist, social democratic, labor parties (with excepti Italian Social Democratic Party) comprise group left parties. Center parties, farmers parties,

500 WORLD POLITICS Summary TABLE 1 Statistics" Variable Mean S.D. Mimum Maximum Total welfare /GDP 23.0 6.2 33.4 11.3 Pensis/GDP 6.9 2.2 3.0 11.0 Public health /GDP 6.0 4.3 0.98.1 Income replacement/gdp 7.1 1.63.4 13.2 Unemployment support/gdp 2.4 1.5 6.6 0.3 Or surance/gdp 4.7 2.3 1.2 9.6 Family benefits/gdp 2.0 0.4 1.24.7 Antipoverty programs/gdp 0.6 03.1 0.6 Housg subsidies/gdp 0.3 01.9 0.4 Inequality (90/10) 1.06 0.25 1.70 0.68 Inequality (90/50) 0.55 0.12 0.88 0.38 Inequality (50/10) 0.51 0.16 0.89 0.27 Right government 41.5 0 36.7 Turnout 78.5 13.2 40.095.6 Percentage elderly 13.5 2.1 9.5 17.7 Unemployment rate 7.2 1.7 3.1 17.2 a See Appendix 2 for data sources. cabet seats held cservative parties each as our measure period partisan compositi government. Summary statistics for our dependent dependent variables are listed Table 1. Fally, it is worth comm discussg ctrols we do not clude. We do not clude measures uni density, uni ccentra ti, or centralizati bargag, sce previous studies have identified se variables as beg primary determants equality wages salaries.33 Our assumpti is effect uni organizati wage-settg stitutis welfare is direct. Unis wage-settg stitutis affect distributi come, which, turn, affects political support for social surance. liberal parties countries with a cservative party right, Christian democratic parties coun tries with a liberal party right, Democratic Party U.S. cstitute group center parties. Cservative parties, liberal parties countries where liberal party is ma party right, Christian democratic parties countries where Christian democratic party is ma party right, plus all small parties furr right comprise group cservative parties. 33 The impact se three variables distributi wages salaries is analyzed Michael Wallerste, "Wage-Settg Institutis Pay Inequality Advanced Industrial Societies," American Journal Political Science 43 (July 1999). For related studies, see Richard B. Freeman, "Labour Market Institutis Ecomic Performance." Ecomic Policy 3 (April 1988); France D. Blau Lawrence M. Kahn, "Internatial Differences Male Wage Inequality: Institutis ver sus Market Forces," Journal 'Political Ecomy 106 (August 1996); David Rueda Jas P tuss, "Wage Inequality Varieties Capitalism," World Politics 52 (April 2000).

EARNINGS INEQUALITY & WELFARE SPENDING 501 The relatiship between organizati labor market wage is so close it is impossible separate effect uni strength per se from effect a more egalitarian wage distributi. We also experimented with ctrols for per capita GDP, trade open ness (imports plus exports as a share GDP), a dummy variable for fed eral systems government as suggested Huber, Rag, a Stephens,34 dummy variable for systems proportial representa ti a measure uni participati government policy forma ti implementati with respect nwage issues developed Traxler, Blaschke, Kittel.35 Ne se variables altered our re sults ccerng all proved be statistically signifi cant most specificatis we tried. Methodological Issues The model we estimate is y it= a +?JV5 + y ' Inequality^ +?'*., + u.t (5) where y. is as a share GDP country i year / = (/ 1985, 1990, 1995), Inequality. t = In (wgo/w10) usg average value m l w9c/wio country from / - 5 / - 1 is vecr ctrol x.t variables. Two methodological issues arise. The first is questi exogeneity our right-h-side variables. The secd ccerns likely deviatis from stard assumptis regardg variances covariances error terms. Two right-h-side variables, particular, might be suspected beg endogenous. Few ecomists would accept assumpti rate unemployment is exogenous with respect un employment benefits. Sce we are not ccerned this article with accurately measurg impact rate unemployment welfare, endogeneity matters unemployment ly extent it alters our ferences regardg y equati 5. Removg 34 Evelyne Huber, Charles Rag, John D. Stephens, "Social Democracy, Christian Democracy, Cstitutial Structure Welfare State,"American Journal'Sociology 99 (November 1993). 35 Franz Traxler, Sabe Blaschke, Bernhard Kittel, Natial Labour Relatis Internatial ized Markets (Oxford: Oxford University Press, 2001). The measure uni participati policy makg with respect nwage issues is described Traxler, Blaschke, Kittel as "associatial (uni) participati state regulati (n-wage issues)" (p. 68). The data are available decade. We assigned 1980-90 figure 1985 our data set, 1991-96 figure 1995 our data set. For 1990, we used average 1980-90 1991-95 figures. We rechecked our results with 1990 removed from our data. In neir case did clusi dex uni participati alter our fdgs with respect.

502 WORLD POLITICS rate unemployment from set ctrols results ly mor changes pot estimates y associated stard errors. Therefore, potential endogeneity rate unemployment does not affect our cclusis regardg welfare. The or variable might be endogenous is our central variable, wages salaries. While ratio is calculated w^ basis pretax wages salaries, welfare system may affect pretax wage salary distributi. Here we rely results Wallerste, who found government had little effect rat*? a^ter w9c/wio ctrollg for uni density, ccentrati uni movement, centralizati bargag, level wage previous period.36 Therefore, we mata assump ti ratio is determed a w9jwl?> country's labor market stitutis is exogenous with respect welfare. The secd ccerns problem implausibility assumpti error terms associated with different countries same year are uncorrelated. The Norwegian government may not csider U.S. a suitable model for its social policy, but Norwegians pay close attenti policy choices made Sweden vice versa. Instead usual assumpti = Eiuii) a2/, a more plausible assumpti is allow for heteroskedasticity cross-sectial correlati er rors. The current cvential approach comparative is politics use OLS obta pot estimates, sce OLS estimates rema unbi ased but correct estimated stard errors for heteroskedasticity cross-sectial correlati.37 However, small sample properties correcti for heteroskedasticity cross-sectial correlati are unclear our data has ly three time periods. To decide wher or not use panel-corrected stard errors, we turned simulatis. The simulatis revealed uncorrected estimates stard errors perform well, even presence heteroskedasticity cross-sectial correlatis, while panel corrected estimates stard errors perform poorly with so few time periods.38 Therefore, we report uncorrected stard errors regressis follow. 36 Wallerste (fn. 33). 37Neal Beck Jathan Katz, "What To Do ( Not Do) with Time-Series Cross-Sectial Data Comparative Politics/' American Political Science Review 89 (September 1995); William H. Greene, Ecometric Analysis, 3rd ed. (Upper Saddle River, N.J.: Prentice Hall, 1997). 38 We generated 400 data sets, each with 45 observatis (15 countries, 3 time periods) 3 re gressors (a cstant plus first 2 regressors column 1 Table 2). The error terms were normally distributed with romly selected, country-specific variances romly selected, cross-natial correlatis. In simulatis 90 percent cfidence tervals ctaed true values co efficients roughly 90 percent time when calculated usg OLS stard errors. When calculated

EARNINGS INEQUALITY & WELFARE SPENDING 503 Table 2 The Impact Inequality Major Categories Welfare Spendg as a Share gdp 18 oecd Countries (1980-95)a 3 4 5 6 Dependent Variable All Welfare Spendg Pensis Health Income Replacement Unemp. Insurance Or Insurance Lagged var. dep. Inequality (90/10) Right govt. Turnout 749*** (.063) -4.50*** (1.50) -.0190*** (.0073) -.0730***.965*** (.056) -0.31 (0.56) -.0051** (.0028) -.0177**.777*** (.103) 0.17 (0.51) -.0047** (.0025) -.0165**.728*** (.065) -3.32*** (0.94) -.0115** (.0044) -.0343**.582*** (.077) -2.12*** (0.48) -.0030 (.0026) -.0141* yrn*** (.064) -1.37** (0.63) -.0070** (.0031) -.0182** Percentage elderly Unemp. rate adj. R2 92.3 (.0250).326** (.170).256*** (.082) 90.7 (.0097).065 (.062) (.0085) -.020 61.2 (.052) 90.4 (.0150).116 (.090).122*** (.050) 82.5 (.0086).163*** (.031) (.0101).116** (.063) -.016 90.7 (.032) > s;.01; **/><;.05; ><s.10 aols estimati; stard errors parensis; n = 50. All regressis clude a cstant. Results We beg with tal welfare as a share GDP. As column 1 Table 2 reveals, tal welfare is significantly negatively related wages salaries. Spendg levels are lower countries are more unequal. Total welfare is also reduced cservative parties government high levels voter turnout. The estimated negative effect turnout social-surance may surprise readers. However, elecrate is both richer older than adult as a populati whole, correlati between elecral participati come is generally weaker than correlati between elecral participati age.39 Thus, lower turnout may an imply older elecrate average. Both share basis panel-corrected stard errors, ctrast, 90 percent cfidence tervals ctaed true value coefficients ly 77 percent time. See Karl O. Moene Michael Waller ste, "Income Inequality Welfare Spendg: Simulatis" (http://faculty-web.at.nwu.edu/polisci/ wallerste/papers.html). 39 Mark N. Frankl, "Elecral Competiti," Lawrence LeDuc, Richard Niemi, Pippa Nor ris, eds., Comparg Democracies: Electis Votg Global Perspective (Thous Oaks, Calif.: Sage Publicatis, 1996).

504 WORLD POLITICS populati who are elderly rate are unemployment positively associated with welfare as a share GDP. Aggregatg all welfare programs ger, however, obscures where how matters. In columns 2, 3, 4 we csider three ma pillars welfare state separately. Each csumes pillar roughly 30 percent tal welfare or 7 percent GDP. In column 2 dependent variable is pensis (old-age cash benefits) as a share GDP. In column 3 dependent variable is government health care as a share GDP. Sce re is little reas thk rate matters unemployment for spend g or pensis health care, estimated coefficient unem ployment is not statistically significant if unemployment is cluded eir we regressi, removed rate unemployment from set ctrols. It is apparent from columns 2 3 has little impact for eir or pensis health care. In both cases, estimated coefficient is not significantly different from zero.40 By ctrast, wages salaries has a significant, negative effect set provide come or replacement surance agast loss come as a result un employment, sickness, disability, occupatial illness or accident, death a spouse (column 4 Table l).41 The estimated impact a permanent crease wage salary e stard de viati (.25) is change come-replacement programs -3.32.25? -0.8 a percent GDP short run (five years) -3.32 -?.25/(1.728) -3.1 percent GDP run. lg Sce average come replacement is 7.1 percent GDP sample, this is a large change. To illustrate with an example, dif ference between average value United ln(w9q/w10) Kg 40 In case health, estimated coefficient is even closer zero if e subtracts means-tested health (Medicaid) from U.S. figures. Excludg U.S. Medicaid (roughly 25 percent tal government health U.S.), column 3 Table 2 becomes yt = 2.92 +.801y^5 +.020 Inequality -.0046 Right -.0123 Turnout -.030 Elderly where stard errors coefficients (excludg cstant) are (.103,.526,.0025,.0088,.052) adjusted R2 = 65.8. Only coefficient lagged dependent variable Right govern ment are at significant.05 level. 41 The category come replacement Table 2 is a subset cluded surance agast loss come Moene Wallerste (fn. 4). The difference between two is measure surance agast come loss Moene Wallerste (fn. 4) cludes a share expen ditures health while all health are excluded from come replacement Table 2.

EARNINGS INEQUALITY & WELFARE SPENDING 505 dorn Sweden early 1990s was.45. That difference wage is estimated be associated with a difference - come replacement 3.32?.45/(1.728) 5.5 percent GDP run. lg The actual difference between come replace ment as a share GDP Sweden U.K. was 7.7 percentage pots 1995 (13.2 percent GDP Sweden as opposed 5.9 per cent GDP U.K.). Thus, difference earngs between United Kgdom Sweden explas about 75 percent actual difference come as a replacement share GDP two countries. The category come-replacement programs can be subdivided provide surance agast risk unemployment, is, sum unemployment benefits active labor-market (2.4 percent GDP average) provide surance agast risks loss come because disabil ity, sickness, occupatial illness jury, death a spouse (4.7 percent GDP average). Results for each se two subcategories come are replacement presented columns 5 6 Table 2. In equality is most strgly related unemployment sur ance active labor-market as, column 5 shows, but relatiship is significant negative for both categories expendi tures.42 It is also worth notg, spite charge employ ers, unis, governments encourage workers apply for disability payments under cditis high unemployment, unemployment rate does not have a significant effect disability surance as a share GDP. In additi, partisan compositi government makes less difference for unemployment surance active labor-market than for any or category social surance. Readers may questi specificati models displayed Table 2. Perhaps unemployment should be dropped from column 6, sce estimated coefficient has "wrg," is, unexpected, 42 An alternative way measure generosity unemployment benefits is replacement ratio, which is available from OECD (fn. 20). Usg average replacement ratio for a worker at median wage at two-thirds median wage first year as unemployment dependent vari able yields yt = 3.81 + - - -.136.864y^ Inequality.0006 Right.0017 Turnout where stard errors coefficients (excludg cstant) are (.048,.054,.0003,.0010) ad justed R2 = 90.5. All coefficients are at significant.05 level. Neir share elderly nor populati rate are unemployment significy different from zero when replacement ratio is dependent variable.

506 WORLD POLITICS sign. Perhaps unemployment rate should be added column 3, sce unemployment may be damagg health. Perhaps cservative government should be removed from set ctrols a priori grounds elecral competiti forces all parties implement same as, Downsian model. Rar than csider each we possible objecti, vestigated robustness results Table 2 regressg each dependent variables lagged de pendent variable, Inequality (90/10), every subset "questi able" ctrol variables, where questiable ctrol variables are Right Government, Turnout, Percent Elderly, Unemployment Rate.43 The results are presented first two columns Table 3, where we display both mimum maximum value estimated coefficient over Inequality all combatis c questiable trols for each dependent variable. Table 3 shows re qualitative sults Table 2 with regard three large compents welfare are budget robust with respect specificati uncertaty. While effect uncertaty regardg correct specificati is larger than samplg uncertaty for any given specificati, every spec ificati implies is negatively associated at.05 sig nificance level with come as a replacement share GDP. In ctrast, is not significantly associated with as a pensis share GDP any specificati. In case govern ment health care, is not significantly associated with as a share GDP most specificatis. To check wher our results could be upset e removg countries from our data set, we redid regressis Table 2 for each subset seventeen countries. The mimum maximum value estimated coefficient Inequality (90/10) are presented third fourth columns?f Table 3. Aga, estimated coefficient Inequality (90/10) is significant at.05 level every subset seventeen countries when dependent variable is tal social surance (le 1), come replacement (le 4), unemployment benefits (le 5) is significant at.05 level all but e subset seventeen countries when de pendent variable is come replacement or than unem 43 This procedure is advocated given a Bayesian justificati Edward E. Learner, Specificati Searches: Ad Hoc Inferences with Nexperimental Data (New York: John Wiley Ss, 1978). We did not csider unemployment rate be "questiable" when dependent variable cluded un employment benefits.

EARNINGS INEQUALITY & WELFARE SPENDING 507 Table 3 The Effect Inequality Expenditures: Robustness Tests51 Extreme Bounds Analysis Jackknife Dependent Variable Mimum Maximum Mimum Maximum 1. All welfare -5.52-2.16-6.90-3.49 (1.50) (1.37) (1.93) (1.61) 2. Pensis -0.75 0.32-0.63 0.20 (0.50) (0.46) (0.59) (0.58) 3. Health 0.10 0.72-0.01 0.50 (0.52) (0.44) (0.55) (0.63) 4. Income -3.46-1.93-4.72-2.54 replacement (0.99) (0.85) (1.19) (0.96) 5. Unemployment -1.83-1.28-1.83-2.50 support (0.68) (0.57) (0.53) (0.51) 6. Or -1.71-0.72-1.79-0.97 surance (0.61) (0.52) (0.86) (0.66) n_50_50_47_47 a Only estimated coefficient for Inequality 90/10 is shown with stard errors parenses. Extreme Bounds Analysis summarizes results 2q regressi equatis all cludg possible subsets q questiable ctrols. Jackknife summarizes re sults eighteen regressi equatis excludg each country e at a time. The Jackknife estimates clude same ctrols as Table 2. ployment benefits (le?).44 In ctrast, estimated coefficient Inequality (90/10) is not significantly different from zero any subset seventeen countries at.10 level when dependent variable is pensis (le 2) or health care (le 3). Krisv, Ldert, McClell distguish between political impact p half wage schedule botm half wage schedule.45 They argue closer median is poor, is, smaller ^5(/w10 wage ratio, greater willgness voters middle support welfare. In ctrast, closer median is rich, is, smaller w^w^ ratio, lower willgness voters middle support welfare. In Table 4 we test proposi 44 In les 1,4,5, 6, mimum estimate is obtaed excludg Austria maximum estimate is obtaed excludg Fl. In le 2, mimum is obtaed excludg Norway maximum is obtaed excludg U.S. In le 3, mimum is obtaed excludg Fl while maximum is obtaed excludg Austria. 45 Lorenzo Krisv, Peter Ldert, Robert McClell, "Pressure Groups Redistributi," Journal Public Ecomics 48 (July 1992).

508 WORLD POLITICS TABLE 4 The Impact 90/50 Ratio 50/10 Ratio Welfare Expenditures as a Share GDPa Dependent 12 3 4 All Welfare Income Unemployment Or Variable Spendg Replacement Support Insurance -4.91-3.60-1.47-2.11 Inequality 90/50 (3.37) (2.02) (1.17) (1.36) -4.19 Inequality -3.13-2.48-0.91 50/10 (2.28) (1.44) (0.77) (0.95) F(l,n-k) 0.20 0.16 0.27 0.29 athe regressi equatis clude all ctrols cluded Table 1 for each dependent variables; n = 50; k = 8 for columns 1,2, 4; k - 7 for column 3. The F statis tic tests null hyposis coefficients Inequality 90/50 Inequality 50/10 are identical. ti 90/50 ratio 50/10 ratio have different political ef fects. The equatis are estimated are identical correspdg equati Table 2, with log o?w^lw^ replaced log w90/^50 log w5q/w1q. Only coefficients In equality (90/50) Inequality (50/10) are displayed. The estimated coefficients always have same sign. Moreover, null hyposis coefficient ln^^/i^) coefficient are same is never ln(^50/^10) rejected. Therefore, our use ln(^90/i?10) = \n{w9jws() + ln(wso/?u10) as measure Table 2 is justified. Pensis, health, come replacement cstitute most, but not all, welfare budget. In Table 5 we present an analysis remag part, divided family benefits services (2 percent GDP average) programs targeted low-come dividuals, refugees, digenous groups plus housg subsidies (1 percent GDP average). Column 1 reveals ne dependent vari ables are good predicrs family benefits, with ex cepti lagged dependent variable. The secd column Table 5 dicates cservative parties are government associated with more housg subsidies antipoverty programs, which probably reflects preference cservative parties for tar narrowly geted programs over broadly targeted programs. In additi, countries with rates high more unemployment spend benefits at targeted those with low come. In neir category, however, is sig nificantly associated with wages salaries.

EARNINGS INEQUALITY & WELFARE SPENDING 509 TABLE 5 The Impact Inequality Smaller Categories Welfare Spendg as a Share gdp (1980-95) 1 2 Dependent Variable Benefits Family Antipoverty Programs Housg Subsidies Lagged dependent variable.521***.986*** (.102) (.075) Inequality (90/10) -0.45-0.24 (0.63) (0.26) Right government -.0025.0028** (.0029) (.0013) Turnout.0150* -.0036 (.0101) (.0044) Percentage elderly.068.024 (.061) (.025) rate Unemployment.035.048*** (.032) (.014) adj. R2_6^8_827_ <; **>.01; <> mp.05; <;.10 *p a OLS estimati; stard errors n = parensis; 50. All stant. regressis clude a c IV Cclusi The empirical relatiship between social-surance as a share GDP advanced dustrial societies differs across. For many?pensis, health care, family benefits, poverty alleviati? is largely uncorrelated with equal ity wages salaries. But for a set significant c stitute roughly 30 percent welfare benefit?unemployment surance, active labor-market, sickness pay, disability sur ance, occupatial illness is jury? significantly more generous countries with a relatively egalitarian pretax distribu ti wages salaries. These differences relatiship between come social-surance across areas can policy be ex explaed tendg Romer Meltzer Richard model corporate fact welfare provide surance as well as redistributi. The dem for redistributi creases when come falls, but dem for redistributive surance creases when come rises. Thus an crease lowers come median voter relative

510 WORLD POLITICS mean generates two counteractg effects. With two counteract g effects, impact support for welfare de pends particular policy under csiderati. Inequality lowers support for provide surance unex agast pected loss come. In welfare where benefits are re ceived all regardless current employment status, two effects roughly balance each or such re is little or no relatiship between come levels. The fact we failed fd any category welfare where clearly raises welfare can be explaed absence social-surance designed purely provide redistributive benefits amg active participants labor market. Our data analysis shows differences empirical rela tiship between earngs different social-surance match predictis extended Romer Meltzer Richards framework. But re are or ex possible planatis we empirical pattern found. Iversen Soskice sug a gest variant surance argument emphasizes relative importance firm-specific skills versus general skills.46 Accordg Iversen Soskice, dem for surance agast job loss is greater countries where firm-specific skills predomate, sce firm specific skills are lost defiti when workers leave ir firm. In countries where general skills predomate, dem for surance agast job loss is less, sce cost job loss is less. In fact, re is a close empirical relatiship between earngs equality mea sures firm-specific skills used Iversen Soskice.47 An egalitar ian wage schedule compresses wage differential between workers at different skill levels creases employers' centive vest firm-specific trag reduces workers' centive vest gen eral trag.48 Thus, effect wage distributi relative importance versus firm-specific general skills is anor route which greater wage equality may crease dem for social sur ance agast job loss. A different approach is emphasize effect wage discentive effects as come-replacement, suggested ^Torben Iversen David Soskice, "An Asset Theory Social Policy Preferences," American Po litical Science Review 95 (December 2001). 47 Iversen Soskice (fh. 46), for example, report a correlati coefficient.73 between ir meas ure extent vocatial trag ratio wxjw^ (p. 889). 48Dar Acemoglu J?rn-Steffen Pischke, "The Structure Wages Investment General Trag," Journal 'PoliticalEcomy 107 (June 1999).

EARNINGS INEQUALITY & WELFARE SPENDING 511 Mfitt, Ribar, Wilhelm.49 They argue if wages at botm come scale are low, n come floor provided social surance benefits must not be so high unskilled workers fd liv g dole preferable workg. The are higher wages at botm, higher come floor provided social surance can be without severe creatg discentive effects. Such discentive effects are less important for publicly provided health surance or pensis. While re is some discussi ways keep elderly work force, voters are much more ccerned about workg-age adults choosg live social-surance benefits stead em seekg ployment than are y about labor-force participati retirees. In our view, political fluence unis, frequently cited as e most important determants cross-natial differences social-surance cannot account, easily for differences exist relatiship between earngs social surance across different categories social surance. We are not aware any evidence suggestg unis care less about pensis health care than about come-replacement programs. After all, retirees make up a significant fracti uni members Europe day.50 Unis crease workers' ability obta workers want, but unis also change workers' over preferences. It is direct effect unis changg workers' prefer ences over social-surance changg distributi come helps expla differential impact across policy categories we have found data. In cclusi, re is more than e reas why social surance agast come loss from layfs, sickness, or accidents might be greater countries with lower levels come. We have emphasized direct impact wage equality political support for redistributive surance agast come loss. Iversen Soskice focus relative importance firm-specific skills versus general skills. Mfitt, Ribar, Wilhelm argue terms dis centive effects come-replacement when wage is high. The negative impact come support for spend g important categories social surance, turn, helps expla strg associati pre-tax transfer come proporti households whose post-tax transfer come falls 49Mfitt, Ribar, Wilhelm (fn. 7). 50 Bernhard Ebbghaus Jelle Visser, Trade Unis Western Grove's Dictiaries, 2000). Europe sce 1945 (New York: