Two decades after Korpi and Palme’s “paradox of redistribution”: What have we learned so far and where do we take it from here?

Abstract Over two decades ago, Korpi and Palme (1998) published one of the most influential papers in the history of social policy discipline, in which they put forward a “paradox of redistribution”: the more countries target welfare resources exclusively at the poor, the less redistribution is actually achieved and the less income inequality and poverty are reduced. The current paper provides a state-of-the-art review of empirical research into that paradox. More specifically, we break down the paradox into seven core assumptions, which together form a causal chain running from institutional design to redistributive outcomes. For each causal assumption, we offer a comprehensive and critical review of the relevant empirical literature, also including a broader range of studies that do not aim to address Korpi and Palme’s paradox per se, but are nevertheless informative about it.


Introduction
Over two decades ago, the Swedish scholars, Walter Korpi and Joakim Palme (hereinafter referred to as K&P) published one of the most influential papers in the history of social policy discipline, in which they put forward a "paradox of redistribution": the more countries target welfare resources exclusively at the poor, the less redistribution is actually achieved and the less income inequality and poverty are reduced (Korpi & Palme, 1998). 1 This paradox has subsequently been a major source of inspiration for many welfare state scholars. Not only is K&P's article one of the most frequently cited in social policy literature, 2 several studies have also stated the explicit objective of revisiting the original analysis on which it was based. K&P's findings remained practically unchallenged for more than a decade but subsequently, a number of comparative studies started re-examining the links between universalism/ selectivism, 3 social spending, and redistribution, taking advantage of more advanced quantitative methods, larger selections of countries, and longer time periods. Among the first to do so were Kenworthy (2011) and Marx, Salanauskaite, and Verbist (2013), who reproduced and updated K&P's empirical analyses, using similar data and methods. Their main conclusions are twofold: (1) the empirical evidence supporting the paradox of redistribution has always been somewhat weak and (2) the paradox might have been valid in the 1980s when K&P carried out their analyses but it certainly no longer applies to the twenty-first century welfare states. Based on the assumption that K&P's argument was "theoretically sound but inadequately operationalized" (Jacques & Noël, 2018, p. 71), later studies moved beyond mere replication and found that by using novel measurements for some of K&P's key conceptssuch as universalismthe paradox of redistribution is far from dead and buried. Brady and Bostic (2015, p. 290), for one, claim that "K&P's conclusions are robust and remain significant two decades later in rich democracies and in a broader sample of developed and developing countries" (see also Jacques and Noël (2018)).
Thus, the scholarly debate surrounding the empirical validity of K&P's paradox of redistribution remains far from settled. Nor has it lost its tremendous societal relevance, especially at present, when increasing inequality and financial insecurity, transforming labour markets, migration and demographic changes exert additional pressures on welfare states and raise important questions about their overall legitimacy and future directions (Taylor-Gooby, Leruth, & Chung, 2017). At one extreme, major international organizations, such as the International Monetary Fund (IMF), EU, Organization for Economic Co-operation and Development (OECD) and World Bank, continue to present means-tested, low-income targeting as one of the most appropriate and promising paths for the future welfare state (Schoukens, 2016). At the other extreme, different types of universal basic income schemes, granting financial assistance to all citizens or residents regardless of their actual needs are being proposed and even experimented within a number of countries (van Parijs & Vanderborght, 2017).
Against this background, the present article contributes to the debate by responding to K&P's call to "open the black box of causal processes assumed to mediate the effects from institutions to redistributive outcomes" (Korpi & Palme, 1998, p. 673). In what follows, we provide a state-of-the-art (though certainly not an exhaustive) review of major empirical research into K&P's paradox of redistribution. More specifically, we break down the paradox into several core assumptions, which together form a causal chain running from institutional design (ie. universalism vs selectivism) to redistributive outcomes (ie. the reduction of poverty and income inequality). For each causal assumption, we provide a comprehensive and critical review of the relevant empirical literature, also including a broader range of studies that do not aim to address K&P's paradox of redistribution per se but are nevertheless informative about it. The main advantage of our approach is that we move beyond a simple test of the "end product" (ie. are selective systems less redistributive than universal ones?) to a much more sophisticated examination of its underlying, in-between mechanisms. While our intention is not to provide the ultimate answer to the central question of the debate, our findings do provide valuable insights that can inform future studies on universal versus selective welfare provision. First, we show that most of K&P's causal assumptions cannot be taken for granted as the relevant empirical evidence tends to be highly conflicting. We also discuss alternative theoretical mechanisms put forward to make sense of the findings that are at odds with K&P assumptions. Second, we identify areas where additional research is very much needed, especially concerning the link between institutional settings, support for redistribution, and the size of redistributive budget. These may be the weakest links in K&P's model as they have not been tested empirically in the original analysis. In addition, the distributional effects of private welfare provision remain under-researched. Third, in practical policy terms, we argue that selectivity can play a role in anti-poverty strategies, but, for this, a number of preconditionssuch as extending coverage to more "deserving" groups like the working poorneed to be fulfilled.
The political economy of universalism and selectivism The fact that distributional justice principles and the methods for selecting beneficiaries can have major implications regarding political support for welfare provisionwhich in turn might affect the programmes' generosity, quality and sustainabilityhas perhaps been acknowledged since the introduction of the Poor Laws in England. Not surprisingly, it has attracted the attention of many prominent scholars, including Besley (1990), Goodin and Le Grand (1987), Korpi (1980), Marshall (1950), Pierson (1994), Sen (1995), Rothstein (1998), Tawney (1952), and Titmuss (1968. In the United States, a major milestone in the debate was the publication of an influential book, The urban underclass in 1991. In this work, two American scientists -Skocpol and Greensteinput forward competing views about which of the two approaches offered more potential for tackling the persistent poverty in the United States. Contrasting the failure of three major targeted programmes (the nineteenth-century poorhouses, pensions for single mothers, and the "Great Society" programmes) with the success of more universal programmes (such as civil war benefits, health education for mothers and children and old-age pensions), Skocpol argues that "when U.S. anti-poverty efforts have featured policies targeted on the poor alone, they have not been politically sustainable, and they have stigmatised and demeaned the poor" (Skocpol, 1991, p. 14). The implications have been a gradual reduction in spending and ultimately the abandonment of the programmes. To avoid the same pitfall, Skocpol urges the advocates of the poor to promote "targeting within universalism": ie. universal programmes in which most people, especially the middle class, would have a direct stake and which would provide disproportionately larger benefits to the disadvantaged.
While not denying that universal programmes tend to have wider electoral appeal and are less stigmatizing than the means-tested alternatives, Greenstein (1991) argues that universalism is not the decisive factor determining the political support for a programme. To substantiate this claim, Greenstein provides a detailed analysis of changes in spending on various social protection programmes during the budget cuts initiated by the Reagan administration, showing that most-targeted programmes fared much better than universal ones, such as disability pensions and unemployment insurance. Based on this evidence, Greenstein argues that targeted programmes can attract substantial political support when, in addition to the very poor, they also incorporate low-income and moderate-income working families, serve more "deserving" groups, such as pregnant women and children, are linked to work and viewed as effective, and do not provide cash benefits to able-bodied people. Thus, instead of "targeting within universalism," Greenstein places greater emphasis on a mixture of universal and means-tested programmes, with a substantial role for targeting, especially given the tight fiscal constraints (and public opposition to higher taxes) prevalent in the United States at the time. Even though the two papers provide sound theoretical reasons for how the underlying distributional principles and other factors might influence the outcomes of welfare programmes, their main limitation is their focus on a single country, which is generally viewed as an exceptional, rather than a typical case of a Western welfare state.
In contrast, K&P sought to approach the issue from a cross-national perspective, focusing on the link between different types of social policy models and aggregate outcomes, such as poverty and inequality. K&P's theoretical starting point is that the Western welfare states differ in institutional structures, as they pursue different "strategies of equality" (1998, p. 664). The targeted model uses a "Robin Hood" strategy, by providing low-level transfers to people on a low income. The basic security model adopts a "simple egalitarian strategy," by giving low-level but equal benefits to everyone, which, in relative terms, is more important for those on low incomes. In contrast, the corporatist model uses a "Biblical Matthew principle," by giving more to the better-off to reward the contributions they have made through making social insurance contributions. The encompassing model also uses this strategy but unlike the corporatist modelwhich excludes non-contributorsit also provides generous support to low-income families. The crucial difference between the three strategies lies in the use of targeting through means testing, which plays a major role in the targeted models, a lesser one in the basic security models, and a minimal one in the corporatist and encompassing models. Another distinction is in the level of transfers, which are low in targeted and basic security models, and high in corporatist and encompassing models.
K&P's central hypothesis, regarding the link between the institutional model and redistributive outcomes to an extent, echoes the one put forward by Skocpol (1991) but entails a more complex causal mechanism that can be broken down into seven core assumptions (see Figure 1). The causal mechanism starts with the prevailing institutional structure of the welfare state, which influences the formation of class coalitions and levels of public support for redistribution (1). Where means testing dominates, the middle-class allies itself with the upper class to contain welfare spending from which it benefits little, driving down public support for redistribution. Where universalism prevails, the middle-class forms coalitions with the working class to alter the market distribution of resources, by redistributing incomes from the upper-income class. The type of class coalitions and levels of support are, in turn, critical for determining the size of the redistributive budget (2). In democracies, sustaining high levels of public spending requires support from a broad coalition of the middle and working classes. For this reason, more universalist welfare states have larger redistributive budgets than selective welfare states (3). The logical next step is that larger volumes of social spending lead to more redistribution (4). In addition, generous welfare benefits reduce the need for private alternatives, which entail no redistribution and would, therefore, lead to less-equal outcomes (5 and 6). Cumulatively, a large volume of spending and limited private provision is the mechanism that explains the "paradox of redistribution": the less the countries resort to targeting through means testing, the more they reduce poverty and inequality. Conversely, the more they try to channel resources to the poor, the less effective they are in achieving their egalitarian objectives (7).
K&P's empirical analysis draws on micro-level data from the Luxembourg Income Study (LIS), while also utilizing several macro-level indicators from the International Labour Organization (ILO) and the OECD. Most of the analysis covers 11 affluent democracies, using LIS data from 1985, and involves a combination of descriptive statistics and bivariate correlations. It shows that more universal welfare states indeed spent a larger proportion of GDP on social protection and had considerably lower rates of inequality and poverty. The relative size of the redistributive budget was found to be strongly positively correlated with redistribution (measured as the difference between Gini coefficients of market and disposable incomes divided by Gini coefficient of market income), while the correlation between the degree of targeting and redistribution was weaker and negative, implying that welfare states relying more on means testing achieve less redistribution. Furthermore, greater selectivity was associated with lower levels of transfers for both social insurance and means-tested benefits. The assumptions concerning the link between institutional design and private provision were tested using the income distribution of pensioners. While public pensions were found to be the most unequal (measured as the difference between the maximum and the average pensions) in the most universalist countries, the opposite was true for pensioners' disposable income. Further analysis confirmed K&P's assumptions, as countries with more equal pensions were found to spend more on private pensions, and, in turn, the spending on private pensions was strongly positively correlated with inequality in the gross income of older people. In sum, the results provided support for most of K&P's assumptions, although it should be stressed that the first two fundamental assumptions concerning the relationships between the institutional design, class coalitions and the size of the redistributive budget were not subjected to empirical testing due to lack of relevant micro data. In the remainder of this paper, we review more recent empirical evidence concerning each of the steps in the causal mechanism.
Reviewing the empirical evidence concerning K&P's causal assumptions Do universal welfare systems generate broader class coalitions and thus greater public support for redistribution?
It has become increasingly clear, however, that it matters a great deal for which specific dependent variables are being studied. When public opinion towards broad and encompassing programmes (such as pensions and healthcare) are scrutinized, no clear-cut regime pattern unfolds, for the simple reason that such programmes are supported by clear majorities in all countries (Bean & Papadakis, 1998;Blekesaune & Quadagno, 2003;Gelissen, 2001a;Gevers et al., 2000;Wendt et al., 2011). Some argue that only popular attitudes towards redistribution and aid to the poor/unemployed, which tend to vary much more across countries, follow a regime logic (Jaeger, 2006;Larsen, 2006;Svallfors, 2003Svallfors, , 2004, at least when attention is limited to some of the closest representatives of Esping-Andersen's worlds of welfare (see eg. Andreβ & Heien, 2001;Linos & West, 2003). In liberal countries, where reliance on means-tested benefits is greater, the support base for redistribution from higher to lower incomes (which is at the core of K&P's argument) does tend to be smaller than in social-democratic and conservative countries, where social benefits are distributed on a more universal basis. The few studies conducted outside of the dominant welfare regime frameworkusing more direct measurements of institutional designseem to confirm that selectivism is indeed associated with lower popular support for redistribution (Brady & Bostic, 2015;Jacques & Noël, 2018).
However, the higher levels of support for redistribution found in universal systems are most definitely not due to the formation of broader class coalitions, as K&P suggested. As far as class conflicts are concerned, the existing body of literature is unequivocal: patterns of attitudinal divisions among different social classes are relatively similar in different worlds of welfare (Blekesaune & Quadagno, 2003;Edlund, 2006;Gelissen, 2000;Jaeger, 2009;Larsen, 2006;Svallfors, 1997Svallfors, , 2003Svallfors, , 2004Svallfors, , 2012Taylor-Gooby, 2001;4 The regime approach to welfare attitudes has always struggled to incorporate some of the additional regime types that have been identified over the years, such as the Mediterranean and post-communist types (Arts & Gelissen, 2002). Expectations regarding the level of support in welfare regimes other than Esping-Andersen's three worlds of welfare have never been clearcut but prior research shows that Southern and Eastern European nations tend to rank highest of all in terms of public support for social welfare (Dallinger, 2010;Svallfors, 2012). Wendt et al., 2011). To the best of our knowledge, only one study suggests otherwise. In revisiting Linos and West (2003) and Svallfors (1997) concluded that class does matter more in liberal regimes (ie. Australia and the United States) than it does in social-democratic (ie. Norway) and conservative (ie. Germany) ones. However, because most other studies have failed to confirm that different regime types create different class cleavages, there is strikingly little evidence that polarization across classes is sharper in selectivist liberal regimes. Furthermore, at the level of specific social policy programmes, there is little clear-cut empirical support for the claim that the middle classes unite with the higher classes in their support for universal benefits and their opposition to means-tested ones (Forma, 1997;Goul-Andersen, 2011;Kangas, 1995;Pettersen, 2001). Thus, even though K&P were correct in arguing that universal systems generate higher support for redistribution, this cannot be explained from the perspective of class interests and coalitions. Instead, others suggest that it has more to do with shortterm self-interest (Pierson, 2001), norms of procedural fairness (Rothstein, 1998) and popular deservingness beliefs (Larsen, 2006). That is, universal systems may yield more public support because they (1) have larger welfare constituencies, irrespective of social classes, (2) are perceived to function better in terms of, eg. administrative cost, complexity and benefit abuse and (3) nurture less-divisive public discussions about the deservingness of welfare recipients (Blomberg & Kroll, 1999;Laenen, 2018;Roosma, van Oorschot, & Gelissen, 2014;Rothstein, 1998;van Oorschot & Roosma, 2017).
Are higher levels of public support for redistribution associated with larger redistributive budgets?
One of K&P's implicit causal assumptions is that broader class coalitions and higher levels of public support for redistribution will ultimately lead to larger redistributive budgets because political elites operating in democratic polities have strong electoral incentives to be responsive to mass opinion (Burstein, 2003;Jacques & Noël, 2018;Page & Shapiro, 1983). Therefore, a certain degree of correspondence between the preferred size of the redistributive budget among the public and the actual size of that budget is to be expected. Only a few studies, however, have examined this opinion-policy link empirically (Myles, 2006). In line with K&P's assumption, Brooks and Manza (2006b, 2006a, 2007 found that both between and within countries, higher levels of public support for redistribution give rise to higher levels of social expenditure (measured as a percentage of GDP). A few years later, Kenworthy (2009, p. 739) replicated Brooks and Manza's empirical analyses but came to the opposite conclusion: "the empirical case for public opinion's influence [on social expenditure] remains to be made." His first argument is that there is no evidence whatsoever for an over-time association between support and spending within countries. The second argument is related to a concern regarding reverse causality, as the "positive cross-country correlation could well be a function of earlier programme generosity causing public support, rather than of public support influencing programme generosity" (Kenworthy, 2009, pp. 732-733), see also Myles (2006)).
Of course, the idea that public policies also shape public opinion is nothing new in comparative welfare state research. In fact, it is one of the central tenets of policy feedback theory and a common assumption in welfare attitudes research (Campbell, 2012;Kumlin & Stadelmann-Steffen, 2014;Larsen, 2006). Most studies within this field treat support for redistribution as an individual-level, dependent variable and social expenditure as a context-level, independent variable. However, the empirical evidence is inconclusive. Without establishing causality, some studies find a positive relationship between social spending and citizens' support for redistribution (Brady & Finnigan, 2014;Dallinger, 2010;Jaeger, 2006) but others report a non-significant association (Jaeger, 2013;Van Oorschot & Meuleman, 2014), or even a negative one (Burgoon, Koster, & van Egmond, 2012;Jaeger, 2013). Much seems to depend on the specific selection of countries, time periods, public opinion measurements, control variables and statistical techniques. To make sense of the contrasting findings, some researchers have argued that the relationship between social spending and support for redistribution is perhaps not linear, as is typically assumed but curvilinear. According to the "saturation" (Dallinger, 2010) or "growth to limits" (Jaeger, 2006) hypothesis, popular demand for redistribution is curbed once the maximum amount of taxes that citizens are willing to pay for the financing of redistributive policies is reached. This should help to explain why Scandinavians, eg. are often found to be less supportive of further redistribution than Southern Europeans, given that they live in mature and already high-spending welfare states (Gelissen, 2001b;Noël & Thérien, 2002). Some empirical studies have used a quadratic term for social spending to capture the curvilinear nature of the spending-support relationship, but have not corroborated it (Dallinger, 2010;Jaeger, 2006).
In sum, the most conservative conclusion that can be drawn from the studies discussed above is that there seems to be a positive, but not necessarily causal, correlation between public support for redistribution and the size of the redistributive budget; a relationship that exists only at the aggregate, between-country level. However, most studies have been so preoccupied with the causality issue that they seem to have forgotten the importance of "needs-adjusting" social spending figures (Clasen & Siegel, 2007), which pose a serious threat to the validity of their findings. Because social expenditure is not only a product of welfare state generosity but also of population composition, the positive correlation with public support for redistribution might be driven by the needs of that population (eg. unemployment rates). Clearly, K&P's causal assumption concerns the link between welfare state generosity and public support, not its relationship to population composition. Therefore, Brady and Bostic (2015, p. 271) use "transfer share," which they define as "the size or extent of the welfare state within the average household's income" and consider "a household-level version of welfare effort" as an alternative measurement for the redistributive budget size. Their results show no significant association between transfer share and public support for redistribution, suggesting that the size of the redistributive budget perhaps does not increase with support levels, after all.
Is the size of the redistributive budget larger in universal welfare systems?
Research concerning the link between the institutional design and redistributive budgets has focused on two related, yet distinctive aspects of the latter: the volume of spending and resilience to retrenchment, either at the (welfare) state-level or that of or specific policies. With regard to the first, studies have sought to verify whether more universalist welfare states are indeed more generous in terms of overall spending. The empirical tests typically involve replicating the method used by K&Pbivariate plots of income redistribution and the volume of spendingas well as modifying the measurements of redistribution and spending in order to better align them with the theoretical framework. In addition, later studies have used more sophisticated regression models that allow several closely related variables to be controlled for, such as GDP per capita, unemployment, and dependency rates (eg. Jacques & Noël, 2018;Jesuit & Mahler, 2004). Several of these studies appear to support K&P's assumption. In a comparative study of income redistribution using 59 LIS data from 13 OECD countries between the 1970s and 2000s, Jesuit and Mahler (2004) report a strong negative correlation between the (low-income) targeting efficiency and the relative size of transfers (the proportion of public transfers in household income). Exploring the effects of taxes and transfers on income redistribution in 24 OECD countries by the mid-2000s, Whiteford (2008) found that countries with systems targeted less at the poor had higher social spending. Similarly, Marx et al. (2013) identified a relatively weak but still positive, relationship between universalism and the generosity of welfare spending in 25 advanced welfare states. In a time-series cross-sectional analysis to revisit the "paradox of redistribution" using data from 2000 to 2011 in 20 OECD countries, Jacques and Noël (2018) found that universalism was associated with higher public social spending as a proportion of GDP. 5 The same authors also identified a positive correlation between universalism and social spending in 21 OECD countries in a more recent paper (Jacques & Noël, 2020).
More ambiguous results were reported by Brady and Bostic (2015, pp. 272-274), who argue that universalism and low-income targeting are not opposite sides of a continuum but represent two separate dimensions; with universalism being "homogeneity across the population in benefits, coverage and eligibility," and low-income targeting a "disproportionate concentration of welfare transfers in low-5 In this study, universalism was operationalized through an index comprising two components: (1) the proportion of social benefits that are means-tested and (2) the proportion of private spending in the total social expenditure. income households." Replicating the "paradox of redistribution" analysis using LIS data with two separate samples of 20 affluent democracies and 37 countries (including those 20 democracies), Brady and Bostic found that universalism was indeed positively correlated with higher spending. However, they also found that low-income targeting was associated with more generous welfare. In the case of the broader sample, they partially explained this seeming contradiction by pointing out that, in some developing countries, social spending was both low and skewed in favour of the higher income groups; once the poor started to be incorporated in the system, both low-income targeting and spending levels improved. However, the positive relationship between low-income targeting and welfare generosity also persisted in the case of the restricted sample. Another study focusing on the level of specific policies, rather than the welfare state level, also provides evidence that disputes the existence of the link between the provision mode and the size of the redistributive budget. In a study of targeting, child benefits, and child poverty in 26 EU countries, Van Lancker and Van Mechelen (2015) found no correlation between the degree to which child benefits were targeted at low-income families and the proportion of GDP allocated for child benefits (including tax credits).
Even if universalism is indeed associated with more generous spending, this in itself does not guarantee immunity from cuts. This is particularly true for mature welfare states, in which there have been different degrees of retrenchment over the last three decades. As a matter of fact, Pierson (1994) pointed out that the large size of the budget makes universal programmes a logical target for politicians seeking to roll back what they see as an overinflated welfare state. Pierson substantiated this argument by tracking the evolution of social protection programmes in the United States and the United Kingdom. during the times of retrenchment initiated by, respectively, Reagan and Thatcher. He found that despite the harsh rhetoric deployed against the beneficiaries of means-tested assistance, the actual retrenchment of these programmes was far less substantial than in the case of universal programmes such as unemployment benefits and child cash transfers. In addition to fiscal considerations, Pierson argues that both administrations also sought to avoid being seen as hostile to the poor; while the public might initially have shared the general criticism of "idle" recipients of social assistance, multiple polling results showed that the electorate did not react positively to cutting the already meagre support for the poor.
The results of comparative studies do not support the idea of selective benefits being more resilient than universal ones, but neither do they lend much support to the opposite hypothesis. In a study using the same sample of countries as K&P, but with longitudinal data for a period between the mid-1980s and the mid-2000s, Kenworthy (2011) found no association between the degree of targeting and changes to redistribution, thus rejecting the idea that more universal systems are more capable of resisting spending cuts. Tracking the retrenchment of general means-tested assistance vis-à-vis three social insurance programmesold-age pensions, unemployment insurance, and sickness insurancein 18 OECD countries between 1990 and 2002, Nelson (2007) found that the social insurance programmes fared better during this period. However, the observed difference between the two types of programmes was relatively small, as both were subject to a considerable decline in generosity when compared with the increases in wages. Moreover, in nearly half of the countries, means-tested assistance was found to be more resilient than at least one of the social insurance programmes. Furthermore, means-tested benefits appeared more vulnerable in more universalist countries than in those that aimed to provide basic security (mainly Anglo-Saxon states). In sum, it appears that more universal welfare states do have larger redistributive budgets, but the institutional design is not a safeguard against cuts.
Does the size of the redistributive budget matter for reducing poverty and income inequality?
The assumption that more spending leads to more poverty reduction appears to be well supported by empirical evidence. Jesuit and Mahler (2004) regressed redistribution achieved through transfers on the size of social budgets, while also controlling for several macro variables, including the progressivity of transfers. They found that social spending had a positive and statistically significant effect. Similarly, using a multilevel model controlling for macro-level variables, Brady and Bostic (2015) found that the volume of social spending had a strong negative effect on relative poverty. The same conclusion can be drawn from Noël's (2018, 2020) studies using regression analysis, as they found strong effects of social spending on poverty and redistribution. Using LIS data from the mid-2000s for a sample of 40 high-and medium-income countries, Ferrarini et al. (2016) also reported that both bivariate correlations and multilevel regression models confirmed higher spending led to a lower incidence of relative poverty among elderly people. Likewise, in a study of single-mother poverty in 18 OECD countries, Brady and Burroway (2012) found that two indicators of welfare generositya welfare state index derived through several welfare spending indicators and the average proportion of median income that households receive in social transfershad strong negative effects on single mothers' likelihood of being poor.
Further supporting evidence comes from a study by Nelson (2004), although instead of directly comparing the effects of budget size (and targeting accuracy), Nelson sought to separate the povertyreducing effects of social transfers into the contributions of means-tested and non-means-tested transfers in five countries: Canada, Germany, Sweden, the United States, and the United Kingdom. The analysis, based on LIS data from the mid-1990s, showed that in all countries, non-means-tested benefits lifted much larger proportions of people out of poverty. Moreover, Sweden and Germanywith much more generous and wider coverage (and thus more expensive) social insurance systemswere considerably more successful in reducing poverty through transfers than the Anglo-Saxon countries.
While the reviewed evidence unequivocally points to the importance of the redistributive budget size for redistribution and poverty reduction, it should be borne in mind that, in some cases, smaller welfare spending can be compensated for by accurate targeting. For example, the study by Whiteford (2008) confirmed the countries that achieved the largest reductions in inequality were universalist ones with large volumes of spending, such as Sweden, Belgium and Denmark. However, the analysis also shows that the United Kingdom achieved the same progress as Germany in reducing inequality, despite spending almost half as much, and Australia achieved the same as France, which spent twice as much. Moreover, zooming in on the proportion of transfers that reached the bottom income quintile, Australia was found to allocate the same volume of cash support (in relative terms) as Norway, despite spending two-thirds of the latter's budget. Similarly, New Zealand allocated as much to the poorest 20 per cent as Luxembourg, despite spending less than half of what the latter did.
Does a larger redistributive budget lead to less private welfare provision, and less private provision to lower levels of poverty and income inequality?
Of all the assumptions in K&P's thesis, the least controversial seems the idea that people are more likely to resort to private provision when they are dissatisfied with public benefits and services. Furthermore, that private provision leads to less-equal outcomes as, unlike social insurance, there is no (or very limited) risk pooling across different income groups. Perhaps, for this reasonbut also because private welfare is a relatively under-researched topic in generalthere is very limited empirical evidence available to test this assumption in a comparative perspective. This is unfortunate, as the scarce evidence points in opposite directions, suggesting that this assumption may also be contentious.
Some evidence in support of both mechanismsprivate social arrangements being a substitute for (inadequate) public welfare provision and private social spending being less redistributiveis provided by a study by Castles and Obinger (2007) into the determinants of social expenditure. In a sample of 18 OECD countries, with detailed data on different types of social spending since the 1990s, the authors identified a negative correlation between (gross) public social spending and (net) current private spending, and the latter was associated with higher levels of poverty and inequality. However, it should be noted that neither of these correlations was statistically significant, while a negative association was shown between higher public spending and inequality and poverty.
More robust evidence corroborating the lower redistributive effects of private spending is offered by Goudswaard and Caminada (2010). In an analysis covering 25 OECD countries, around 2005, they tested the correlation between redistribution and public and private social spending, with the types of spending expressed separately in gross, as well as net (after tax) terms. The analysis showed that higher public spending was associated with more redistribution, while higher private spending was correlated with less redistribution. However, when disaggregating the analysis at the programme level, Goudswaard and Caminada also found that, while higher spending on all public programmes except incapacity benefits was associated with more progress in reducing inequality, within private arrangements, only higher spending on pensions was correlated with less success in redistribution.
Some work, however, provides evidence that indirectly contests K&P's assumptions. In a study covering 14 OECD countries using LIS data, Marier and Skinner (2008) found that stronger reliance on private pension provision was not associated with larger gaps between women's and men's pensions. In a smaller sample of nine countries, they also found no evidence that more generous minimum pensions reduce the need to invest in private pension schemes. In addition, Gelepithis (2016) pointed out that the link between the generosity of public pensions and private provision may work in a way that differs from that suggested by K&P: instead of prompting people to opt for private alternatives, means-tested pensions can reduce the incentives for building up private savings. This, she argues, explains why several OECD countries, especially the Anglo-Saxon ones -Australia, Ireland, New Zealand and the United Kingdomhave, since the 1990s, consistently tried to expand the coverage of contributory and universal schemes, in order to limit means testing and the associated disincentives for saving.
Do universal welfare systems redistribute income and reduce poverty more than selective systems? Kenworthy's (2011) study was one of the first that produced results contradicting K&P's ultimate causal link between institutional design and welfare outcomes. While confirming the strong correlation between redistribution and the level of universalism in 1985, Kenworthy found that it gradually weakened before disappearing completely by 2005, as some "universal" countries became more selective (Denmark), while "selective" ones became more universal due to the maturation of old-age pension systems (eg. the United States). To explain this phenomenon, Kenworthy suggested a refinement of K&P's thesis. He speculated that universalism matters at the initial stage of welfare state development: universal social insurance programmes assure the middle classes that the welfare state is primarily a social risk management tool rather than a vehicle for redistribution from the rich to the poor. In this context, it is easier for policymakers to target other programmes to people on low incomes, as these programmes will be viewed as of minor importance. Kenworthy argued that this is why Danish policymakers were able to resort to more selectivity without compromising the volume of spending and redistribution. On the other hand, the lack of similar large-scale universal programmes did not foster strong support for welfare provision among the American middle class; therefore, an increase in the total spending did not translate into more redistribution.
Further evidence of the weakening or disappearance of the link between universalism and redistribution was offered by Marx et al. (2013Marx et al. ( , 2016, who explicitly sought to replicate the results of K&P's and Kenworthy's studies, as well as extending the analysis to a sample of 25 affluent countries. Particularly valuable in Marx and colleagues' studies is the use of multiple redistributions (the difference between Gini coefficients before and after tax and transfers versus the difference between market and gross incomes), income rank measurements (gross income vs. disposable income versus market income), and the alternative treatment of mandatory private and occupational pensions (market income vs. social transfers). This provides solid robustness checks of their key findings. Depending on the methodological choice, Marx and colleagues found that, in the early 2000s, there was either no association or a very weak positive one between low-income targeting and redistribution. Furthermore, decomposing the achieved redistribution by benefit types, they found that social assistance benefits made a disproportionately large contribution to the reduction of inequalities compared with other benefits, and that all highly redistributive systems contained strongly selective subsystems. Marx and colleagues' explanation for this echoes Kenworthy's hypothesis that the overall universality of a system allows the introduction of selective transfers without much public resentment. However, they also emphasized the changing role of selectivity in advanced welfare states: whereas in the mid-1980s social assistance was associated with work disincentives and dependency culture, its subsequent extension to low-income employed people and linking to the activation agenda have made it more popular both among the electorate and policymakers.
Equivocal results were reported by Brady and Bostic (2015). On the one hand, in contrast to K&P, they found that low-income targeting was negatively associated with relative poverty in both restricted (20 democracies) and extended (37 democratic and non-democratic countries) samples. However, once they simultaneously controlled for the transfer share and for universalism (defined as homogeneity in transfers, coverage, and eligibility, as noted above), the association between low-income targeting and poverty was no longer robust. Universalism was also negatively correlated with poverty in both samples, but after controlling for low-income targeting and transfer share this association was no longer statistically significant. Overall, while corroborating some of the key assumptions of K&P's paradox, Brady and Bostic suggested two new paradoxes of social policy based on the relationships they identified between poverty, support for redistribution, and their measurements of universalism and low-income targeting. The "non-complementarity" paradox implies that there is a mismatch between the determinants of poverty and support for redistribution: on the one hand, a higher transfer share reduces poverty but it is not related to support for redistribution; on the other hand, low-income targeting reduces support for redistribution but is not related to poverty. A positive correlation between the transfer share and low-income targeting leads to the second, "undermining" paradox: a higher transfer share implies stronger low-income targeting, which leads to less support for redistribution, thus undermining the poverty-reducing efforts achieved by the higher transfer share. Jacques and Noël (2018) found a positive, bivariate correlation between universalism and poverty reduction, although, as noted before, this disappeared when they controlled for social spending in a regression model. Thus, they suggested that the effects of universalism on poverty are mainly indirect, as proposed by K&P: more universalism leads to higher social expenditure, which in turn reduces poverty. Based on this, Jacques and Noel argue that K&P's thesis is theoretically correct and continues to be valid in the first decade of the twenty-first century. However, they suggest that K&P's operationalization of universalism was not appropriate, as it was measured by the concentration of transfers, which is influenced by factors other than low-income targeting, such as pre-transfer income distribution and the demographic profile of welfare beneficiaries. For example, if families with children are more likely to be poor than others, a universal benefit for them will appear as targeted towards low-income groups. Jacques and Noel proposed that this misspecification can be avoided if universalism is measured through policy design instead of redistribution outcomes. Hence the rationale for the measurement of universalism they used in their analysis; an index comprising two indicators: (1) the proportion of social benefits that are means tested and (2) the proportion of private spending in the total social expenditure.
In a later study, using an almost identical sample, Jacques and Noël (2020) again reaffirmed the superiority of universalism but they added important qualifications to their argument. Within universalism, they propose a critical difference is made by the balance between old-age pensions and benefits for working-age population. Old-age pensions, especially those with strong second tier (contributory element), tend to favour the better-off, whereas benefits for working-age people are more pro-poor even without means testing due to the vulnerable situation of the recipients (ie. unemployed, families with many children, disabled people). Therefore, with a similar level of spending, those universal welfare states that direct a higher proportion of total spending towards working-age population reduce poverty more effectively than those that prioritize old-age pensions more.
Similar to Jacques and Noël (2018), Van Lancker and Van Mechelen (2015) also argued that assessments of policy intentions are more appropriate to capture the extent of universalism of a system than measurements of policy outcomes derived through transfer concentration indices. Thus, they constructed a targeting indicator measuring the variation in average child benefits that families with different income levels were entitled to. On this basis, Van Lancker and Van Mechelen found that, contrary to K&P's findings, a higher degree of selectivism was associated with a greater reduction in child poverty, mainly because in more selective systems, low-income families received more generous benefits. However, they also stressed that the size of the redistributive budget matters, and that the countries in which the largest progress in reducing child poverty was achieved were those that used targeting within universalism, ie., universal systems with higher benefits for low-income groups (see also Skocpol, 1991).

Conclusion
While certainly not the first to explore the political economy of universal and selective welfare provision, K&P's investigation made an immense contribution to the study of the subject by developing a sophisticated multi-stage causal model and testing it empirically in a comparative context. What is more, like any exceptional study, it has inspired a large body of research, which with an increasing availability of high-quality comparative data and more advanced methods has provided a more nuanced understanding of the distributional implications of the institutional setup of Western welfare states.
However, 20 years after K&P's initial study, there are substantial reasons to cast doubt on the validity of the causal model, even though the original empirical analysis has withstood rigorous re-examinations. As our review shows, the only assumption, ie., unequivocally supported by more recent studies is that higher welfare spending is associated with lower poverty and inequality but even in this regard there is some indication that countries can compensate for lower spending by more accurate targeting of low-income families. Furthermore, there is ample evidence that the very first stages in the causal mechanisminstitutional structures of the welfare state influencing the formation of class coalitions and the latter affecting the size of the redistributive budgetare simply not correct, as class coalitions do not differ between welfare regimes. Perhaps, one of the main drawbacks of K&P's study is that the very foundations of the causal mechanism were taken for granted based on theoretical reasoning instead of empirical material.
There are several other reasons, both objective and subjective, for the assumptions and the overall thesis possibly not being valid at present. One is that Western welfare states have changed substantially in the past 20 years. To deal with the persistent problem of "in-work poverty," governments have extended the coverage of selective programmes to also incorporate people in paid work but on low incomes, who attract more sympathy from the public (and policy-makers) than the "idle" poor (Marx et al., 2013). At the same time, governments in most Western countries have sought to foster private provision, eg. by encouraging and mandating participation in private pension schemes (Ebbinghaus, 2015). These changes have undoubtedly affected the mechanisms in K&P's causal chain. In addition, strong emphasis on activation policies and tougher conditionality placed on the recipients of social assistance (and unemployment benefits) may have also softened their image in the eyes of the wider public.
Other issues are the data limitations and methodological and theoretical choices made by K&P. A sample of 11 countries is far too small to draw conclusions that can be generalized to all welfare states, especially when using the relatively basic method of bivariate correlations. For example, if the original sample had included Mediterranean welfare states, which combine low levels of selectivity with low levels of redistribution, and are also characterized by a great deal of occupational segmentation in their social insurance systems, the observed associations would have undoubtedly become far weaker. 6 Furthermore, while it is often used to measure the progressivity of social transfers, a concentration index is a somewhat imperfect measurement of redistribution, as it is also influenced by the demographic profile of the sample and the pre-transfer distribution of incomes. Similarly, it has been shown that when the focus is shifted from gross to net social spending to take into account the fact that some countries tax certain social transfers while others do not, some countries (especially the universalist Scandinavian ones) are not as high spenders as is commonly assumed (Adema, Fron, & Ladaique, 2011).
Despite these substantial causes for scepticism about the validity of K&P's causal chain, the fact remains that more universalist countries tend to have better redistributional outcomes. In this respect, perhaps we should bear in mind the caution expressed by K&P (1998, p. 673) that "Institutional structures, however, are only one of many factors that affect the distribution of income in a country … Political traditions, demographic composition, labour force participation rates, levels of unemployment, wage setting practices, and industrial structures are also important." In addition to these matters, there is a strong rationale for considering a confounding variable problem suggested by Marx et al. (2013, p. 8): An extensive welfare state, as well as a limited degree of wage inequality may both be the result of variables that determine both… A society in which the value of solidarity is widely shared may simultaneously support pay norms, collective agreements and adequate minimum wages, as well as quasi-universal and generous benefits.
Beyond the purely academic value, the reviewed evidence has practical implications that can be taken on board by both policymakers and those who advocate on behalf of the poor. While any general recommendation risks the fallacy of one-size-fits-all, it seems reasonable to suggest that in their pure forms neither universalism nor selectivism can be the most optimal remedy for addressing poverty and inequality. Targeting, however, seems to work best when embedded within the overall framework of universalism (Marx et al., 2013(Marx et al., , 2016Van Lancker & Van Mechelen, 2015). Hence the rationale for a combined approach of deploying selective measures alongside and within universal programmes. The required balance between universalism and selectivism will inevitably vary across countries but in any context, it seems important that selective policies are not too narrowly targeted at the very poor. Instead, they should also include low-income groups; especially those viewed as more deserving, either because they lack control over their situation (eg. children), or because they pay back society in some way (eg. people who work in low-paid jobs) (van Oorschot, 2006). In this respect, working tax credits and wage supplements seem particularly promising. In parallel, other ways of improving targeting of lowincome people within the universal framework, eg. by redirecting expenditure from old-age pensions towards certain categorical working-age benefits as proposed by Jacques and Noël (2020), also appear promising and deserve further consideration.
Before concluding, we would like to point out several potential avenues for further research in addition to the need for considering the confounding variable issue and a more comprehensive set of structural determinants of income distribution, as argued above. There is surprisingly limited empirical evidence on the social legitimacy of universal vis-à-vis selective welfare provision. Future studies could narrow this gap by analysing the primary divisions (ie. socio-demographic and ideological) in these preferences and how (and why) these vary across countries and over time. Similarly, research into private welfare provision has not kept pace with the actual policy developments in this field. In addition to analysing the distributional implications of increased reliance on private provision, future research could also shed light on the drivers of the general trend: whether it is inadequate public benefits that push people towards private alternatives or whether it is the high costs of universal provision that prompts policymakers to encourage (and mandate) the privatization of social risks. Lastly, it should be stressed that with the notable exception of Brady and Bostic (2015) and Ferrarini et al. (2016), almost all of the available evidence comes from developed welfare states. Extending the geographic coverage to lessdeveloped countries could help to determine how universal the political economy of universalism versus selectivism is.