Literature review

Almost every advanced democracy has some form of minimum income protection (MIP). In their Handbook of Minimum Income Protection in Europe, Bahle et al. define MIP as ‘a social minimum based on a means test’ (Reference Bahle, Hubl and Pfeifer2011: 13). The two elements are important in this definition: the reference to a social minimum means that minimum income protection very much defines a social protection floor for citizens, an income below which nobody should fall; and the means test element specifies the manner in which this income is attributed, not as a universal right or as a function of past contributions, but as a last resort for those below a certain income (usually with little or no assets, and no family support). For unemployed and uninsured working-age adults, the MIP basically corresponds to social assistance and related transfers.

For many years, it was considered impossible to compare minimum incomes across countries, because social assistance regimes appeared as a maze of categorical programmes with arcane rules, abundant exceptions, local variations and numerous in-kind advantages. In a landmark report released in 1996, Tony Eardley and his co-authors used the family-type approach to estimate in purchasing power parities the social assistance benefits offered to various types of households by the different OECD countries. With data for 1992 only, they established important differences in minimum income protection across countries, with Switzerland, the Netherlands, the Nordic countries and Australia standing among the most generous (Eardley et al., Reference Eardley, Bradshaw, Ditch, Gough and Whiteford1996: 137). Their conclusions, however, pointed to numerous variations in rules and regulations, which led them to classify countries in a range of categories, each one containing only a few cases. It appeared difficult to rank countries along a single monetary continuum, let alone connect them to characteristics of the broader welfare state.

A year later, Ivar Lødemel published a thorough comparative analysis of social assistance policies in Norway and Britain, which also concluded that the general configuration of the welfare state was a poor predictor of social assistance arrangements (1997). Lødemel even identified a welfare paradox whereby social assistance generosity appeared contrary to a country's welfare state regime. In Norway's universalist welfare state, social insurance was so extensive that social assistance became ‘relegated to obscurity and expected to vanish,’ and remained anchored in the Poor Law tradition. In Britain's liberal welfare state, on the contrary, the prevalence of means-tested programmes gave importance to social assistance and made the ‘alleviation of poverty’ a ‘main social policy objective’ (1997: 261–66). Hence, the paradox: the most encompassing welfare state became less engaged in the expansion and modernisation of social assistance.

The first effort to test these conclusions comparatively came with the early work of Nelson. In his Ph.D. dissertation, he paired the minimum income protection data of Eardley et al., with the welfare state indicators gathered in Stockholm for the Social Citizenship Indicators Program (SCIP). Using qualitative comparative analysis (QCA) and ordinary least squares (OLS) regression, he found, contrary to Lødemel, a positive relationship between the protection offered to the middle class by social insurance programmes and the generosity of minimum income protection, confirming an important implication of the power resources theory of the welfare state (Nelson, Reference Nelson2003: 125). There were limits, however, to what could be concluded from a set of 18 country cases at one point in time. In later works, Nelson developed his dataset to cover the years 1990–2002, and used time-series cross-sectional analysis to assess the connection between welfare state regimes and minimum income protection. It did not seem possible, it turned out, ‘to distinguish any clear cross-national patterns or groups of countries that correspond neatly with previous attempts to cluster welfare states into certain institutional types’ (Nelson, Reference Nelson2008: 114). These findings, Nelson contended, did ‘not at least contradict previous claims about potential institutional relationships between minimum income protection policies and first-tier benefits, such as social insurance’, but the relationship remained difficult to establish. In later works, Nelson found much diversity in minimum income protection, but no clear correspondence between income adequacy and types of welfare states (Montanari et al., Reference Montanari, Nelson and Palme2008).

In a dissertation presented in 2009 at the University of Antwerp, Natascha Van Mechelen took up Nelson's initial project of elucidating the socio-economic determinants of minimum income protection. Using both fuzzy set and time-series cross-sectional analysis, she found no clear relationship between benefits generosity and socio-economic conditions such as the government's financial liabilities, the unemployment rate or the proportion of social assistance recipients, for an average of benefits estimated for four types of households. Like Nelson in his later work, Van Mechelen failed to corroborate a relationship between the broader welfare state, measured by the structure of social insurance programmes, and MIP adequacy (Reference Van Mechelen2009: 164). She found, however, that strong trade unions proved favourable to social assistance recipients, a conclusion consistent with power resources theory (2009: 189–91). Van Mechelen's findings, like those of Nelson before her, suggested that the adequacy of minimum income protection resisted the neat theories or classifications that prevailed in the study of the welfare state.

The early findings of Eardley and his collaborators about the complex and sui generis character of social assistance programmes thus seemed to stand (1996). As Lødemel suggested, the general configuration of the welfare state appeared to be a poor predictor of social assistance arrangements (1997). In a recent book on minimum income protection, Ive Marx, Kenneth Nelson and their co-authors focus on general trends and do not address determinants (Reference Marx, Nelson, Marx and Nelson2013). ‘Most large-scale comparative investigations on social assistance and other forms of minimum income benefits’, concur Birnbaum et al. in a new book, ‘[remain. . .] descriptive in nature, with little or no reference to political dynamics or wider policy contexts’ (Reference Birnbaum, Ferrarini, Nelson and Palme2017: 61).

Theory

To account for the lack of connection between social insurance programmes and the adequacy of minimum income protection, Van Mechelen refers to the Matthew effect argument, whereby mainstream programmes benefit primarily the middle-class and barely reach the poor (2009: 130). Universal programmes may boost middle class support for social spending, as the power resources approach suggests, but this support does not necessarily extend to programmes aimed at the poor. In fact, programmes designed for the middle class could even crowd out income support for the poor (2009: 131–32). There is, of course, a classical reply to this Matthew effect/crowding-out argument, best articulated by Walter Korpi and Joakim Palme in their ‘Paradox of Redistribution’ article (Reference Korpi and Palme1998). Encompassing insurance programmes, they argue, broaden the basis of support for welfare state expansion and enlarge the size of the redistributive budget, two factors that facilitate improvements in income support for the poor (1998: 672). Van Mechelen acknowledges this possibility but notes that a number of studies have challenged the validity of the paradox of redistribution argument (Van Mechelen, Reference Van Mechelen2009: 140; for recent demonstrations along these lines, see Kenworthy, Reference Kenworthy2011; Brady and Bostic, Reference Brady and Bostic2015; Marx et al., Reference Marx, Salanauskaite and Verbist2016). At best, the connection between the welfare state and minimum income protection appears uncertain.

This connection, however, may simply have been inadequately assessed. There are indeed good theoretical reasons to think there are links between the welfare state and minimum income protection. First, the institutional configuration argument at the heart of welfare state literature remains highly plausible. Second, the related political argument drawn from the power resources approach also seems credible, and it is partially validated by Van Mechelen's own findings on the impact of union density.

Consider, first, the institutional argument. As Esping-Andersen argued, the very purpose of the early fights for the welfare state was to provide citizens with ‘a socially acceptable standard of living independently of market participation’. Decommodification, he contended, was ‘the alpha and omega’ of welfare state politics (Reference Esping-Andersen1990: 37). Social assistance incomes may not be the top priority of socialist parties and trade unions but these incomes were inevitably pushed up by the decommodification of social relations. In addition, as Korpi and Palme explained, the ‘size of the budget available for redistribution is not fixed’ and it is greatly conditioned by the institutional characteristics of the welfare state (1998: 663). An encompassing welfare state should thus have a positive impact on MIP adequacy in two ways: through the logic of decommodification, which makes generous social assistance benefits more acceptable, and through the size of the redistribution budget, which makes generous social assistance benefits more feasible.Footnote ³ Two hypotheses capture these welfare state effects:

Beyond institutional determinants, the politics of the welfare state should also influence MIP adequacy. In this case, two contrary arguments appear plausible. The first line of reasoning stems from power resources theory, which presents politics as the democratic expression of class conflicts, where workers and their allies are represented by parties of the left and trade unions favourable to generous social insurance and transfer programmes (Van Kersbergen and Vis, Reference Van Kersbergen and Vis2014: 48–50). In this perspective, the predominance of leftist parties in government and a high level of union density should favour MIP adequacy. David Rueda raises doubts about this argument and contends instead that leftist parties and trade unions mostly defend the interest of labour market insiders, and neglect the preferences of outsiders for passive income protection (2007: 68). If this is true, the power resources argument should not apply to minimum income protection, social assistance recipients being the perfect outsiders. Rueda's standpoint, however, may be overly rationalistic. In his study on the politics of poverty, David Brady acknowledges that leftist electoral coalitions are primarily composed of voters ‘not vulnerable to falling into poverty’ and rarely make poverty alleviation a core objective (2009: 103). These coalitions, however, are bound by ‘ideological motivations’ that cannot be reduced to ‘strict material interests’. They constitute what Brady calls ‘latent coalitions for egalitarianism’, broadly favourable to redistribution (Reference Brady2009: 103–104). In his own work, Brady finds the cumulative power of the left and union density to be negatively related to poverty (109–15). Qualitative evidence also suggests that leftist governments are more likely to improve benefits for those who do not work (Larocque and Noël, Reference Larocque and Noël2014). Hence, even though MIP adequacy may never be a top priority of leftist parties and trade unions, we can nevertheless hypothesize, in line with power resources theory, that the power of the left influences social assistance incomes:

These effects may be weaker than those of welfare institutions and, as Brady finds for poverty, they may be mediated by institutional variables. They should nevertheless play a role, in the expected direction.

Budgetary constraints should also matter. In her work on minimum income protection, Van Mechelen considers the influence of the size of the public debt as a proportion of GDP, the unemployment rate and the social assistance rate; the latter two measuring indirectly the demand for public support. She finds no significant relationship for the public debt variable, and weak relationships for unemployment (negative) and the social assistance rate (positive). In this article, we consider only the public debt and unemployment variables because there are no reliable OECD data on the social assistance rate. In any case, the unemployment and the social assistance rates are likely to be strongly correlated. With respect to budgetary constraints, we thus have two hypotheses:

Trends and measures of association

Between 1990 and 2012, the adequacy of minimum income protection diminished almost everywhere. On average, as can be seen in Figure SM1 (in the online supplementary material), minimum incomes were relatively stable until 1995, not too far below 50 per cent of the national median income. They then decreased for ten years, to fall below 40 per cent by 2005. The average adequacy then stabilized, slightly above the 40 per cent of median income level (Marx et al., Reference Marx, Nolan and Olivera2014: 27). This decline of adequacy (12.9 per cent in 23 years) was largely due to the growth of median income or, more precisely, to the fact that benefits failed to follow the evolution of median income. Indeed, between the mid-1980s or early-1990s and the mid- or late-2000s, the real average annual growth in median household income for these countries (minus Switzerland) stood around 1.8 per cent (own calculations, based on the national averages provided in OECD, 2011: 43).

This common downward trend did not necessarily imply a convergence among countries. If anything, as Figure SM2 (online supplementary material) on the evolution of the coefficient of variation indicates, variation among countries increased over the years.

Most countries, however, moved in the same direction and became less generous. Figure 2 displays national trends over time. Starting and arriving points are different, and the pace and timing of change vary, but most countries went from better to worst adequacy, except perhaps Germany, Italy and the Netherlands. Ireland also stood out, with a U-shaped curve of regression and recovery.

Figure 2.

Adequacy of minimum income protection, 18 OECD countries, 1990–2012.

Consider, now, the bivariate relationships between the different variables, when data for the 18 countries are pooled over 20 years. The results are presented in Table 1.

TABLE 1.

Correlations between the different variables, 18 OECD countries, 1990–2010 (with p value and number of observations).

Note: ***significant at 0.01 level.

As expected, there are positive relationships between welfare state institutions and the level of adequacy. The index of generosity is strongly related to adequacy and so is the measure for social expenditures. The second set of hypotheses, about power resources, also seems validated. The left's cumulative power and union density are positively associated with adequacy, contrary to the idea of a social-democratic or union indifference to labour market outsiders. The strength of the left and of trade unions is also positively related to the index of generosity and the level of social expenditures, a finding in line with the literature (Van Kersbergen and Vis, Reference Van Kersbergen and Vis2014: 48–50). As expected, there is a negative relationship between adequacy and what is probably the main economic constraint, public debt as a percentage of GDP. The relationship with unemployment is negative as well, but not significant. Our hypothesis on unemployment as a constraint can be dropped (H_3c).

To go beyond bivariate relationships and understand how the different variables interact, it is necessary to build a multivariate model. With a small number of cases over 20 years, the best approach consists in pooling country cases and years in a panel analysis, with a time-series cross-sectional regression model. For this model, we consider the variables that are significant in Table 1: namely, welfare state generosity, social expenditures, cumulative left power, union density and public debt. The general form of the model is:

$$\begin{eqnarray*} \hspace*{-5pt}&&{\textit MIP}\;{\textit Adequacy} = a + {b_1}\;{\it Welfare}\;State\;Generosity + {b_2}\;Social\;{\it Expenditures}\\ \hspace*{-5pt}&&+ {b_3}\;{\it Cumulative\;Left\;Power} + {b_4}\;{\textit Union\;Density} + {b_5}\;{\textit Public\;Debt } \end{eqnarray*}$$

Time-series cross-sectional analysis

There are vigorous debates among scholars using time-series cross-sectional models over the merits of various strategies. One approach, long dominant in political science, consists in pooling all cases together, ignoring the heterogeneity among clusters (countries in our case). This is the approach proposed by Nathaniel Beck and Jonathan Katz with their Panel Corrected Standard Errors (PCSE) procedure (Reference Beck and Katz1995). This strategy has the advantage of being simple but it ignores unobserved heterogeneity and may induce an omitted variable bias. More precisely, it obscures the fact that variations between and within countries may not be alike (Bartels, Reference Bartels and Franzese2015). Such differences in variations are usually important and many scholars prefer to rely on a fixed effects (FE) model, which controls out the heterogeneity between cases to focus solely on change within them. Because it makes abstraction of the national context and solves the problem of unobserved variations, the fixed effects approach has become more or less the ‘gold standard’ in econometrics and political science (Bell and Jones, Reference Bell and Jones2015: 139). These models have problems of their own, however. Indeed, they make it impossible to assess the impact of time-invariant or slowly changing variables, often important in political science (Beck and Katz, Reference Beck and Katz2001; Plümper et al., Reference Plümper, Troeger and Manow2005; Bell and Jones, Reference Bell and Jones2015: 139). Welfare state generosity, trade union density, or the public debt do not move a lot in a given country over 10 or 20 years, but they may make a lot of difference across countries. Because it controls out variations across cases, the fixed-effects approach disqualifies at the outset the very type of variables that comparative political scientists consider most important. A third approach consists in using random effects (RE) models, or a ‘partial pooling’ approach which estimates a weighted average of the variations between and within cases. Theoretically satisfying, this approach also poses technical challenges, because time-varying variables risk being correlated with the random effects term, a problem that usually leads scholars back to fixed effects models, with their own limitations (Bartels, Reference Bartels and Franzese2015).

In recent contributions, Brandon Bartels (Reference Bartels and Franzese2015) and Andrew Bell and Kelvyn Jones (Reference Bell and Jones2015) propose to address this dilemma by modeling explicitly the between-cases and within-cases effects, to avoid confounding them (as in the complete or partial pooling approaches) or ignoring half of them (with fixed effects models). The idea is to generate cluster-specific variables to allow the modeling of distinct effects. We may expect, for instance, that the effects of political and institutional variables will be significant between cases, whereas budgetary constraints and social expenditures will be influential within cases. Because it is a new approach to time-series cross-sectional analysis, let us start by considering the results we would obtain with PCSE or with conventional fixed-effect models.

For our PCSE model, we follow the approach defended by Stephens and his collaborators, and consider a model without lagged dependent variable or fixed effects, with a Prais-Winsten correction (AR(1)) to correct for first-order auto-regression without misestimating other variables (Huo et al., Reference Huo, Nelson and Stephens2008; Huber and Stephens, Reference Huber and Stephens2012: 136). Our other FE models are a conventional robust regression with fixed effects and a fixed-effect model with AR(1) correction (Wilson and Butler, Reference Wilson and Butler2007). In all these instances, the New Zealand case is dropped for lack of data on public debt.

The results of the PCSE model, presented in Table SM2 (online supplementary material), seem to validate our main hypotheses. Welfare state institutions (H_1a), social expenditures (H_1b), and trade union density (H_2b) have a significant positive impact, and public debt as a percentage of GDP (H_3a) a negative impact on MIP adequacy. The left's cumulative power, however, is not significant (H_2a). Many of these variables are slow-moving institutional characteristics unlikely to remain significant in fixed effects models. This is precisely what happens with the two FE models presented in Table SM2 (online supplementary material), where only social expenditures remain significant in both models, in the expected direction (H_1b).

To clarify the relative importance of slowly moving and more incremental variables, the best option, as explained above, is to disentangle variations between and within cases. Table 2 presents results for a model constructed along the lines suggested by Bartels.Footnote ⁴ This model is a robust regression and includes a lagged dependent variable to account for dynamics (Bartels, Reference Bartels and Franzese2015).

TABLE 2.

Random-effect model separating between-country and within-country effects for the determinants of MIP adequacy, 17 OECD countries, 1990–2010.

Robust standard errors in parentheses.

***p < 0.01, **p < 0.05, *p < 0.1.

Results for this random effect, between-within model are consistent with those of the PCSE model, but carry more information about the specific effects of long-term and short-term variables. Indeed, it becomes clear with this model that the index of generosity explains differences between countries, whereas social expenditures account for change within cases. Political variables, however, fare less well. Trade union density is positively associated with adequacy across cases but it is not significant (p = 0.115). The cumulative power of the left is significant within cases but, surprisingly, the effect is negative, suggesting that countries where the left was successful in the 1990s and 2000s became less generous with social assistance incomes. It may be, as Jane Gingrich and Silja Häusermann suggest, that in recent years social-democratic parties have found themselves ‘under pressure to move away from policies supporting traditional income replacement and decommodification and towards new middle-class reform priorities’ (Reference Gingrich and Häusermann2015: 55). Before we reach this conclusion, however, it would be prudent to consider the importance of the observed effect. Finally, public debt works against adequacy, both between and within countries.

Since there is no national social assistance programme in Italy, we also tested the model without this country. We did the same with the United States, a clear outlier with respect to minimum income protection. The model was also tested for the 2000s only. The results, which are reported in the online supplementary material (Table SM3), proved consistent with those of the full model.

Standardised coefficients, presented in Figure 3, help make sense of these different effects by providing a standard measure of impact.

Figure 3.

Standardised coefficients, random-effect model separating between-country and within-country effects for the determinants of MIP adequacy, 17 OECD countries, 1990–2010.

As can be seen in Figure 3, the most important effects on MIP adequacy are cross-sectional, with welfare state generosity and public debt as a proportion of GDP both exerting a strong influence. For one standard deviation in the generosity index, MIP adequacy goes up by about 0.57 standard deviations; for one standard deviation in public debt, adequacy is lowered by 0.40 standard deviations. By comparison, within effects appear modest. A one standard increase in social spending causes a 0.10 increase in adequacy, and a similar increase in public debt lowers adequacy by 0.03 standard deviations. The negative impact of the left's cumulative impact must be interpreted in this light: an increase of one standard deviation in leftcum70 only lowers adequacy by 0.07 standard deviations.

Substantively, these findings suggest that national differences matter more than change across time, which is not so surprising given the relative stability of MIP adequacy. A country's type of welfare state and comparative level of public debt weight more than its changing levels of social expenditures, cumulative left power, and public debt; three variables that are nevertheless significant to explain variations within countries. Overall, the outcome seems more consistent with the logic of decommodification identified by Esping-Andersen (Reference Esping-Andersen1990) than with Lødemel's welfare paradox (Reference Lødemel1997), and it lends support to hypothesis H_1a. Our second welfare state hypothesis, H_1b, about the size of the redistributive budget, also appears vindicated. The effect, in this case, plays over time. As predicted by Korpi and Palme (Reference Korpi and Palme1998), generous social insurance programmes favour the poor. These conclusions corroborate Nelson's early results (Reference Nelson2003), which were not confirmed by Nelson and Van Mechelen's later work (Nelson, Reference Nelson2008; Van Mechelen, Reference Van Mechelen2009). They also confirm the institutional and power resources theories, rather than the Matthew effect and crowding out hypotheses. All in all, for social assistance recipients, it appears preferable to live in an encompassing welfare state with a sizeable redistribution budget and a public debt under control.

The politics of social assistance appears trickier. In line with the power resources approach, with Van Mechelen's findings (Reference Van Mechelen2009), and with our fourth hypothesis (H_2b), trade union density is a bivariate correlate of MIP adequacy but the effect disappears in a multivariate model including institutional variables. Contrary to the insiders/outsiders theory (Rueda, Reference Rueda2007), trade unions seem to participate in what Brady calls a ‘latent coalition for egalitarianism’ (Reference Brady2009: 102). Parties of the left, however, do not have the same impact. If anything, the consequences of social-democratic cumulative power are marginally negative, at least within countries, forcing us to reject hypothesis H_2a. Perhaps, as Gingrich and Haüsermann suggest, social-democratic parties have moved away from traditional income-replacement policies (Reference Gingrich and Häusermann2015). Or, as Bea Cantillon suspects, their preference for social-investment policies worked against the poor, unable to take advantage of programmes meant to facilitate job integration and make work pay (Reference Cantillon, Cantillon and Vandenbroucke2014). In a recent article, Nelson concurs with this interpretation, and suggests public expenditures on active labour market policies (ALMP) have been negatively correlated with benefits adequacy (Reference Nelson2013: 393–4). Such a trade-off between activation and minimum income protection, however, is far from evident. In the late 1990s and early 2000s, the leftist politics of the Third Way often combined an emphasis on activation with a commitment to redistribute (Huo, Reference Huo2009; Larocque and Noël, Reference Larocque and Noël2014). One should keep in mind that the negative effect of leftcum70 is very small and that simple correlations between cumulative left power and MIP adequacy remain positive and rather strong. It may simply be, as Brady suggests, that the effects of leftist politics, like that of trade union density, disappear because they are ‘channeled through welfare generosity’ (Reference Brady2009: 114). Nevertheless, as Brady acknowledges, leftist electoral victories do not seem to bear promises of immediate improvements for the poorest (114).

Finally, as we expected, economic constraints matter. Contrary to the results obtained by Van Mechelen (Reference Van Mechelen2009), public debt as a proportion of GDP has a significant and negative impact on minimum income protection, and this is true within as well as across countries, lending strong support to hypothesis H_3b. The general idea that financial constraints place the welfare state under stress seems vindicated (Van Kersbergen and Vis, Reference Van Kersbergen and Vis2014; Marchal et al., Reference Marchal, Marx and Van Mechelen2014).