Income Risk and Social Spending: Empirical Estimates

Castronova, Edward Article Income Risk and Social Spending: Empirical Estimates Schmollers Jahrbuch – Zeitschrift für Wirtschafts- und Sozialwissenschaften. Journal of Applied Social Science Studies Provided in Cooperation with: Duncker & Humblot, Berlin Suggested Citation: Castronova, Edward (2002) : Income Risk and Social Spending: Empirical Estimates, Schmollers Jahrbuch – Zeitschrift für Wirtschafts- und Sozialwissenschaften. Journal of Applied Social Science Studies, ISSN 1865-5742, Duncker & Humblot, Berlin, Vol. 122, Iss. 3, pp. 327-349, https://doi.org/10.3790/schm.122.3.327 This Version is available at: https://hdl.handle.net/10419/292019 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by/4.0/ Schmollers Jahrbuch 122 (2002), 327-349 Duncker & Humblot, Berlin Income Risk and Social Spending: Empirical Estimates By Edward Castronova* Abstract This paper uses panel data on developed countries to estimate simultaneous equations models of social spending. The methods take advantage of some recent innovations in the growth literature involving the treatment of country-level panel data. Another contribution is to treat income risk as an endogenous variable, as suggested by the recent theoretical work of Hans-Werner Sinn. The results indicate that social spending is moderately related to aggregate income variability, and strongly related to the share of elderly and unemployed. Zusammenfassung In diesem Beitrag werden - basierend auf Panel-Daten aus Industrieländern - mit Hilfe simultaner Modelle die Sozialausgaben geschätzt. Diese Methoden beruhen auf einigen neueren Erkenntnissen in der Wachstumstheorie, welche Panel-Daten auf Länderebene mit einbeziehen. Ein anderer Beitrag dieses Papiers besteht darin, das Einkommensrisiko als endogene Variable zu behandeln, wie dies in einer kürzlich von Hans-Werner Sinn veröffentlichten Arbeit vorgeschlagen wurde. Die Ergebnisse zeigen, dass die Höhe der Sozialausgaben in geringem Maße von den aggregierten Einkommensvariabilitäten abhängt und in starkem Maße mit der Höhe des Anteils von Älteren und Arbeitslosen korreliert. JEL-Classification: H5,13 1. What explains the level of social spending? This paper uses a database of developed countries from 1960 to 1994 to assess the impact of a number of country-level variables on social spending. The two main contributions of the paper are 1) to apply practices recently developed in the growth literature to the question of social spending determination, and 2) to pay serious attention to the role of income risk as a causal factor in social spending. * The author would like to thank three anonymous referees for their comments on the paper. Schmollers Jahrbuch 122 (2002) 3 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/schm.122.3.327 | Generated on 2023-04-04 12:30:24 328 Edward Castronova The literature on this subject, broadly speaking, is an effort to determine why the Welfare State is a more important thing in some countries than in others. The origins and causes of social spending (which most authors take as the best empirical proxy for "The Welfare State") are undoubtedly complex and there are many theories about them, some of them formal, many of them not. There is, however, no formal theory that is both general enough to encompass a significant number of different motivations, and that yields estimable equations, so a rigorous modeling and estimation strategy is not possible. As a result, empirical research on the overall size of the Welfare State has adopted a different strategy. First, general concepts are laid out in order to identify the kinds of variables that ought to covary at the country level. Then basic regressions are estimated in order to test the predicted relationships. There are many such approaches in the literature; Esping-Anderson (1990), for example, traces the causes of the Welfare State to generalized historical 'worlds' or mind-sets involving the degree of conservatism, labor market institutions, and devotion to free-market capitalism. He then argues for this grouping on the basis of a large number of simply-specified OLS regressions on OECD cross-sections. Pampel and Williamson (1989) propose informal theories based on class, voting groups, institutions, politics, and macroeconomic indicators; their supporting evidence comes from a series of fairly basic GLS regressions. Similarly, Hicks and Swank (1992) assume that social spending is driven mostly by the structure of the political process and national institutions and run regressions of social spending on a series of political variables. Contributors in Flora and Heidenheimer (1981) also look for sources of the Welfare State in the intensity of left politics and general economic conditions. This research is most successful in laying out general notions of Welfare State motivations.1 Still, none of these papers make use of what is now a very large literature in public choice economics in which social spending is traced to the rational decisions of individual agents. Specifically, many have argued that income risk is an important determinant of the political demand for social insurance.2 Hans-Werner Sinn (1996) has developed an empirically tractable formal theoretical version of this argument, which remains largely unexplored in multiple-equation empirical work (although see Bird, 2001; Katzenstein, 1985; Cameron, 1978). 1 See also: Uusitalo (1984), Hicks and Misra (1993), Huber, Ragin, and Stephens (1993), Baldwin (1990). 2 The argument has been made in conceptual work (Barr, 1992; Esping-Andersen, 1990) and historical treatments (Rimlinger, 1971; Baldwin, 1990). Schmollers Jahrbuch 122 (2002) 3 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/schm.122.3.327 | Generated on 2023-04-04 12:30:24 Income Risk and Social Spending: Empirical Estimates 329 The empirical methods in this literature, moreover, suffer from serious problems: there is little attention paid to problems of causality, and the presence of unobserved fixed effects is ignored. As for the first issue, in most of these simple regression approaches, social spending is treated as the only endogenous variable, with everything else in society assumed to be an exogenous causal force. In reality, social spending is co-determined with other important social conditions, including the level of income, the degree of income risk, the amount of private investment, and possibly even the degree of inequality. When endogenous variables are treated as exogenous, coefficients reflect a simple correlation only; they indicate neither the size nor direction of causation from the RHS variable to the dependent variable. As for the fixed effects problem, ignoring fixed effects in country-level data is effectively to ignore the existence of unobserved history: unobservable events, institutions, and forces that have an impact on the dependent variable. Ignoring them can lead to a misinterpretation of historical correlations as true causal forces. For example, the US has had historically higher levels of risk and historically lower levels of social spending than countries such as France and Sweden. Unless all of these historical differences are accounted for by some variable in the data set, a simple cross-country regression using contemporary data will lead one to conclude that the correlation between spending and risk is negative. It may be the case, however, that in all the countries and at all times within the current data set, an increase in social spending from its historical norm will lead to an increase in risk from its historical norm. Thus, the cross-country pattern endowed to the data set by history suggests a negative correlation, but the causal flow is actually positive. As a result, ignoring fixed effects can lead to biased conclusions about contemporary influences. Some papers do take account of these problems, and this paper will pick up where they leave off. Specifically, Peter Lindert uses methods that account for the endogeneity problem. He has two papers on the level of social spending, one using a remarkably extensive data set from 1880-1930 (Lindert 1994), and another with a more contemporary data set from 1960 to 1981 (Lindert, 1996). He successfully estimates models in which social spending is jointly determined with income growth. Focusing on a political pressure-group theory of social spending, Lindert finds that democracy, demography (i.e. age-group sizes), and the income distribution have the most influence. Surprisingly, the deadweight costs of social spending are found to have little impact on growth.3 3 Some of the most intriguing Lindert results can be explained through the risk framework. The finding that the Welfare State does not reduce growth can be explained by the fact that the income insurance effect of the Welfare State encourages risk-taking and thereby growth (Bird, 2000). There is a result that spending falls as Schmollers Jahrbuch 122 (2002) 3 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/schm.122.3.327 | Generated on 2023-04-04 12:30:24 330 Edward Castronova This paper will adopt Lindert's multiple equations approach, and will add to it in three ways. First, the database here will be somewhat larger, in terms of years, countries, and variables. Second, this paper will consider income risk as an important, and endogenous, determinant of social spending. Third, this paper will account for fixed effects in the determination of social spending. To summarize the literature, there has been broad interest in determining the empirical causes of social spending at the national level, but more work can be done on the details and rigor of the empirical modeling, as well as adding new concepts such as income risk. A strategy of formal modeling and testing still seems impractical, because the Welfare State has too many complex explanations to synthesize in a single testable model. Nevertheless, more rigorous analytical attention can be paid to the way that even broad and informal theories of social spending translate into specific empirical implications. 2. Conceptual underpinnings: risk and social spending In any such exercise, it is necessary to discuss at least briefly the kinds of variables that are thought to have some kind of influence on the size of the Welfare State in a given country. Fortunately, the previous literature has suggested a number of possible determinants, including inequality and poverty, the population share of politically powerful and entitled groups, and political measures.4 For example, even though it would be open to considerable debate, most authors assume that the best metric for the Welfare State is the share of social spending in GDP, and we will follow that convention here. To focus attention on a relatively unexamined argument for social spending (at least in terms of formal empirics), consider Hans-Werner Sinn's (1995) argument for the importance of risk preferences in determining the size of the national budget. In essence, Sinn claims that a polity that enjoys investment and entrepreneurial activity may call on its government to increase social spending as a hedge against the risks that these activities entail. In Sinn's model, a country is using the Welfare State as a tool to help it the gap between the middle income and lower incomes rises, which might be explained as follows: the middle class assesses its own risks of poverty by the distance between its incomes and those of the poor. As this gap widens, the perceived risk falls, so the demand for income-insuring social spending falls. 4 For a limited overview of some of the conceptual arguments for the Welfare State, see Trattner, 1999; Himmelfarb, 1992; Bird 1999; Hochman and Rogers (1969); Becker (1985); Kristov, Lindert and McClelland (1992); Meltzer and Richard (1981); Piven and Cloward (1971); Mead (1997). Schmollers Jahrbuch 122 (2002) 3 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/schm.122.3.327 | Generated on 2023-04-04 12:30:24 Income Risk and Social Spending: Empirical Estimates 331 choose its desired bundle of national risk and national income. There is some empirical evidence that the income insuring, anti-risk, effects of social spending can be substantial, even for middle class households (Bird, 1995, 2000). Other authors have pointed out that the middle class is often a strong supporter of the Welfare State (LeGrand, 1987; Pierson, 1996), and some argue that this is largely a self-insurance motivation (Atkinson, 1995). If indeed the Welfare State is a form of income insurance, it may actually promote economic growth. Sinn's ideas will be used as the basic framework for the empirical work. 3. Empirical methods To implement Sinn's (1995) ideas about risk and social spending, consider the following simultaneous equation model of social spending, risk taking, and income generation. The unit of observation is a country-year; let yit denote the income level in country i in year t, rit the level of income risk (i.e. the variance), and sit the level of social spending (empirical definitions of these variables in the data at hand will be given below). Each of these three variables is endogenous: Vit = ayi + Pyrfit + PysSit + Py^Xyit + PyA^it + £yit fit = ari + PryVit + PrsSit + Pr&rit + £rit Sit = <*si + PsyVit + PsrTit + PsZ^sit + esit where the terms refer to exogenous variables, k is a measure of the (endogenous) capital stock or investment level, the a and (3 terms are parameters, and the e terms are random errors. The intercepts are country-speci- fic, which will call for a fixed-effects estimation strategy. The first equation is a fairly standard aggregate income equation, familiar from the growth literature (Temple, 1999). The important parameters are /3yr , which measures the presumably positive impact of risk-taking on the income level, and /3ys, which shows how social spending directly affects income. If deadweight costs are substantial, this should be negative. In the second equation, ¡3^ indicates the impact of higher incomes on the willingness to take risks; if r and y are defined in levels, declining absolute risk aversion would imply Pry > 0. The other risk coefficient, /3rs measures the impact of social spending on risk - if the Welfare State encourages risk-taking, then f3rs > 0. Thus while pys shows the direct impact of the state on incomes, and presumably is dominated by deadweight costs, (3rs and 0yr show an indirect and presumably positive effect: the state encourages risk-taking, and risk-taking encourages growth. In the spending equation, f3sy measures the reaction of so- Schmollers Jahrbuch 122 (2002) 3 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/schm.122.3.327 | Generated on 2023-04-04 12:30:24 332 Edward Castronova cial spending to income; if the Welfare State is a normal good, f3sy > 0. Finally, psr shows how spending responds to the risk level. If voters facing higher risks are inclined to call for more income insurance (as the risk motive suggests they should), then 0sr > 0. The structural coefficients in the model already provide useful information about the validity of the risk motive for the Welfare State. To provide information on the other motives, key exogenous variables will be added to the social spending equation (and the other equations where it seems theoretically appropriate). Anyone working with country-level data will be familiar with the difficulties that arise when making judgments as to which variables are exogenous and which are endogenous. Suffice it to say that this paper seeks primarily to extend previous research by adding risk as an endogenous variable. Making other variables endogenous (such as human capital and investment) would require still more identifying variables and thus put an even greater strain on the data. At a broader level, it will always be necessary to make some judgments as to which variables are endogenous and which are exogenous; otherwise, it is simply not be possible to say anything about causation. However, recall that most of the existing literature makes no effort at all to sort out causal effects. Single-equation OLS is the norm. This paper attempts to explore the causes of social spending by applying a structural model to the data, which requires making some assumptions about exogeneity; whether or not these are good assumptions, the fact remains that even making the effort here is an advance on the existing literature. The exogenous variables are chosen based on conceptual arguments in prior literature, and will be limited to some extent by the data. They include: - the gini coefficient (which will also be treated as an endogenous variable, and a mismeasured variable) and a measure of infant mortality; - measures of the size of entitled voting-age populations, such as the aged and the unemployed, as well as a unionization score to account for the political power of labor; - data on the number of strike days lost, the extent of military expenditures, the vote share of left parties, and the degree of voter turnout. By altering how these variables are defined and used, it should be possible to get a sense of which relationships are robust in the data. There are some serious limitations in the kinds of variables that can be used, however, because of the difficulty of finding comparable cross-national data. These limitations and other aspects of the data will be discussed in the next section. Assuming that the data come in the form of a panel of countries over several years, the data can be transformed by calculating the time average of Schmollers Jahrbuch 122 (2002) 3 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/schm.122.3.327 | Generated on 2023-04-04 12:30:24 Income Risk and Social Spending: Empirical Estimates 333 the dependent variable, y, and then subtract it from yit for each observation. Applying the same process to the right-hand side of the regression equations sweeps out the fixed effect terms ay , ar, and as . In effect, this ensures that the parameter estimates will take account of (and be unbiased by) any country-specific factor that is constant throughout the time frame of the panel. This would include any historical forces, observable or not, whose effects occurred before 1960. Thus, historical differences in inequality, civil liberties, war experience, political culture, and religious traditions are all fully accounted for in these results. All of these forces have created historically normal levels for all of the variables, from which each variable in the data evolves from 1960 onward in the course of the panel. What is measured here is the effect of this contemporary evolution of independent variables on the contemporary evolution of the dependent variables. In fully accounting for history, then, these results provide the most accurate estimate of contemporary causality.5 With the transformed data, it is then possible to apply standard 2SLS techniques to the simultaneous equations. The coefficients will be identified if each of the endogenous variables (at a minimum y, r, s, and k) is instrumented by exogenous variables that do not appear in the other equations. As Temple (1999) has pointed out, with panel data on countries, each endogenous variable can be instrumented by its own lags. This seems to be an almost unavoidable choice, since the possibility of finding good instruments among contemporary variables - in other words, national aggregates that one is forced to assume do affect one thing currently but not some other thing currently - would seem to be slim. It is much more plausible that past values of a current variable do affect it strongly but do not have a strong affect on other current variables. In what follows, then, all endogenous variables are instrumented with three lag variables, in addition to other exogenous variables where exclusion seems plausible. The subject of lagged variables brings up another methodological wrinkle: with national aggregate data, how does one account for the fact that years may pass before a shock to one variable has its causal influence on another? One approach is to apply and then explicitly analyze the pattern of lag effects, but this is needlessly complicated (especially so in a multipleequation system with multiple lags). A simpler response is to define all the 5 An alternative method for achieving the same results would be to assume that the historical norm effects a are not fixed parameters, like the a terms, but are unobserved random variables in the error term. This leads to random effects regression, which is mechanically not very different from fixed effects (Greene, 1993, pp. 466- 71). Conceptually, the fixed effects assumption makes more sense here. The data here consist of a census of the available population (countries), each with a fixed history; this is not random sample from a large population where each observation has an unobservable individual shock term. Schmollers Jahrbuch 122 (2002) 3 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/schm.122.3.327 | Generated on 2023-04-04 12:30:24 334 Edward Castronova variables as averages within fixed time windows. Thus, effects are not interpreted as the instant effect of a shock to one variable upon another, but as the sustained effect on one variable of a sustained change in the other. For example, because of political bottlenecks and implementation problems, it is unlikely that an increase in income risk will immediately cause an increase in social insurance, even if the pressure for change is present. However, if the risk shock is sustained for, say, three years, spending has time to respond. To allow for these lagged effects, then, all variables are expressed as three-year averages. In sum, we will estimate three equations of deviations-from-the-mean, where both the mean and the deviations are constructed from variables that are defined as averages over three-year time windows. Variables that are endogenous variables are instrumented by lagged values, and estimates are obtained using 2SLS. Because of the distinct possibility that the error terms in the equations might be correlated within countries, the reported standard errors are huber-white robust standard errors with clustering by country. As it turns out, taking account of clustering reveals that the usual standard errors are strongly biased downward. 4. Data The study makes use of country-year panel data and is collected from two sources. The main source of data is the Comparative Welfare States (CWS) data set, compiled by Evelyne Huber, Charles Ragin, and John D. Stephens (Huber, Ragin, and Stephens 1997). The CWS contains comparable country time series from 1960 to 1994 for 19 developed countries (including Australia, Austria, Belgium, Canada, Denmark, Finland, France, Germany, Ireland, Italy, Japan, Luxembourg, Netherlands, Norway, New Zealand, Sweden, Switzerland, the UK, and the US). Together these countries constitute virtually a census of the developed western world in the post-WWII period, including representatives from all of the major Welfare State 'models' (scandanavian, conservative-corporatist, laissez-faire), as well as other countries (Japan, Ireland) that do not fit these nice categories. The CWS provides a wealth of social spending categories as well as demographic data, macroeconomic data (including a subset of data from the Penn World Tables), and political data. The initial source for most of the series used in the paper is either the ILO or the OECD. The results also make use of some unionization data compiled by Jelle Visser (Visser, 1996). In addition to the CWS, the paper makes use of the Deininger and Squire compilation of inequality estimates (Deininger and Squire, 1996). The Deininger and Squire data have recently been subjected to criticism Schmollers Jahrbuch 122 (2002) 3 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/schm.122.3.327 | Generated on 2023-04-04 12:30:24 Income Risk and Social Spending: Empirical Estimates 341 and it does seem to have a powerful positive effect (beta = 0.379). The demographics of recipient populations, such as the aged and unemployed, should increase spending, and this is the case. These two coefficients are universally large, positive, and statistically significant throughout. Conversely, indicators of social stress such as strikes (beta = -0.290) and turnout (beta = - 0.051), as well as left vote shares (beta = -0.041) are small and have unintuitive signs. The signs vary, but the coefficients only rarely appear to be large or statistically significant.8 As for anti-poverty motives for total social spending, the evidence seems to argue against it: the coefficient on the gini is large, negative, and statistically significant throughout virtually all the results. This runs directly counter to the view that social spending is mainly driven by a desire to help the poor. Infant mortality has a positive but not large or statistically significant impact, and is not robust. As already mentioned, left voting has a negative impact. While some of this evidence could be considered only inconclusive, the general trend seems to argue against compassion for the poor as a main motive for social spending. Table 3 repeats the estimation using the income growth rate in place of the income level in the income equation. From a welfarist perspective, it would seem that the income level, which determines utility, would be of greater interest than the growth rate, which has only indirect implications for well-being. Still, most of the literature focuses on income growth rather than levels, so this table is included to allow comparison to the literature. Most of the patterns from Table 2 are repeated, in particular those of greatest interest here, in the social spending equation. One difference there is that growth has a negative impact on spending. This is consistent with a convergence theory of growth: the smaller the country, the higher the growth rate. Hence if social spending is lower in poorer countries (see Table 2), then it should be lower where growth rates are highest. Another difference worth noting is in the growth equation, where now social spending apparently deters growth (beta = -0.188). This is again conceivable through a convergence theory: social spending makes countries richer (Table 2) but richer countries do not grow as quickly Setting aside the convergence idea, however, the question of whether or not the Welfare State imposes a significant drag on the economy depends on one's object of interest: well-being or development. It seems to raise well-being but slow the rate of development. (The result is robust across multiple variations in methods, not shown.) Note, however, that the gini coefficient has no noticeable impact on growth or the income level, and in both tables the gini reduces social spending as well. This pat- 8 The political literature (Flora and Heidenheimer, 1981; Hicks and Misra, 1993) actually has not been able to establish clearly that left governments have a larger effect on social spending. Schmollers Jahrbuch 122 (2002) 3 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/schm.122.3.327 | Generated on 2023-04-04 12:30:24 342 Edward Castronova Table 3 Growth regression results Dependent Variables Independent 1. Income 2. Risk 3. Social Spending Variables beta s.e. beta s.e. beta s.e. Growth - - .028 .022 -.453 *.066 Risk .096 .329 - - .545 .472 Social Spending -.188 *.070 .024 .033 - - Investment .150 *.078 - - - - Inflation -.294 *.038 .035 *.013 -.162 *.044 Unemployment .007 .128 .023 .026 .291 *.085 Aged - - - - .871 *.189 Trade .064 *.020 -9.3e-4 .004 .070 *.020 Union .089 *.017 -.003 .007 .044 .034 Gini .002 .017 .008 .006 -.089 *.037 Workers .077 .082 - - - - Capital -.151 *.036 - - - - Strikes -.025 .641 -.074 .249 -.255 .874 FLFPR --.010 .009 -- Kids - - 1.949 *1.038 - - Mortality - - - - .069 .094 Turnout -- - - -.028 .039 Military - - - - .108 .440 Left - - - - -.051 *.024 R2 .4531 .1225 .7882 N 344 344 344 Note: coefficients identified with a '*' are statistically significant at the 90 percent level, twotailed test. Source: Comparative Welfare States data set; Deininger and Squire inequality data. tern runs counter to that predicted by a set of recent theories on the role of inequality in growth, which argue that inequality discourages growth because it causes social spending, which is a growth deterrent (Persson and Tabellini, 1994, Aghion, Caroli, and Garcia-Penalosa, 1999).9 9 That literature (see Persson and Tabellini, 1994; Aghion, Caroli, and Garcia-Pe- nalosa, 1999) uses a slightly different method, regressing subsequent growth rates on some initial inequality measure in a reduced-form model. That is, the procedure is not to regress social spending on inequality and then growth on social spending, as is done here. Instead, growth is directly regressed on inequality, with the results that inequality at the start of some time period causes lower growth in later years. Here the finding is slightly different: contemporary innovations in inequality have no apparent effect on contemporary innovations in growth, either directly or through the mechanism of social transfer. The difference in methods is probably dictated mostly by a difference in data; the Comparative Welfare States data base allows examination of social spending, but is limited to developed countries, while the Heston-Summers Penn World Tables do not have social spending but allow examination of developing countries. Schmollers Jahrbuch 122 (2002) 3 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/schm.122.3.327 | Generated on 2023-04-04 12:30:24 Income Risk and Social Spending: Empirical Estimates 343 Tables 4-6 return to income in levels and focuses on the social spending equation, with other relevent coefficients included as an addendum at the bottom of the table. Table 4 presents results based on a different approach to estimating risk. In the base case, risk is estimated from a GARCH model on log income, in regression 4.1, it is estimated from a GARCH model on the Table 4 Variations in risk and income definitions Dependent Variable: Social Spending Independent Variables 1. Risk calculated from income levels, not logs 2. Permanent income (from GARCH) replaces observed income; risk based on levels, not logs 3. Income in logs, not levels; risk based on log income beta s.e. beta s.e. beta s.e. Income .174 .150 .400 *.155 5.676 •1.402 Risk -.662 .453 -.500 .439 .793 .486 Inflation -.022 .041 -.029 .041 -.039 .042 Unemployment .377 *.103 .384 M01 .370 M03 Aged .870 *.247 .675 *.265 .702 *.223 Trade .049 *.029 .053 *.027 .050 *.029 Union .030 .038 .038 .037 .028 .035 Gini -.096 *.042 -.089 *.040 -.102 *.041 Strikes -.624 .882 -.558 .845 -.365 .828 Mortality .103 .114 .133 .108 .141 .116 Turnout -.053 .042 -.046 .042 -.057 .039 Military .369 .505 .362 .505 .305 .486 Left -.040 .032 -.039 .031 -.034 .029 R2 0.7534 0.7597 0.7590 N 344 344 344 Addendum: Coefficient on risk in income equation -.668 *.341 -.470 .316 -.016 .027 Coefficient on social spending in income equation .122 *.074 .161 *.064 .017 *.008 Coefficient on social spending in risk equation .021 .018 .028 .022 .041 .026 Note: coefficients identified with a '*' are statistically significant at the 90 percent level, twotailed test. Source: Comparative Welfare States data set; Deininger and Squire inequality data. Schmollers Jahrbuch 122 (2002) 3 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/schm.122.3.327 | Generated on 2023-04-04 12:30:24 344 Edward Castronova Table 5 Variations in risk and social spending definitions Dependent Variable: Social Spending Independent Variables 1. Risk defined as squared deviations from trend of log income (no GARCH) 2. Social spending defined as spending per capita (not relative to GDP), in 000 3. Social spending defined as nonhealth social spending relative to GDP beta s.e. beta s.e. beta s.e. Income .310 M62 .218 *.047 .453 .305 Risk .152 .198 .142 *.046 .515 .349 Inflation -.032 .042 -.011 .009 -.041 .048 Unemployment .339 M10 .052 *.002 .497 *.160 Aged .728 •.247 .171 *.062 .983 *.416 Trade .056 *.027 .007 .004 .066 *.035 Union .007 .045 .011 *.007 .094 *.051 Gini -.102 *.040 -.023 *.007 -.159 *.049 Strikes -.200 .824 .343 M22 2.000 *.756 Mortality .039 .110 .026 .022 .223 *.133 Turnout -.052 .047 -.003 .010 -.036 .048 Military .492 .496 -.041 .069 .233 .452 Left -.043 .032 .004 .008 -.023 .063 R2 0.7391 0.8345 0.6787 N 319 311 311 Addendum: Coefficient on risk in income equation .030 .106 .118 .184 .317 .291 Coefficient on social spending in income equation .168 *.092 1.2e-3 *3.9e-4 10.415 6.981 Coefficient on social spending in risk equation .047 .066 1.9e-3 *8.4e-4 1.36 1.30 Note: coefficients identified with a '*' are statistically significant at the 90 percent level, twotailed test. Source: Comparative Welfare States data set; Deininger and Squire inequality data. income level, and then expressed as a fraction of income. There is no substantial change in the social spending pattern, but risk now has a large, negative, and statistically significant impact on income, a direct contradiction of the insurance motive. In regression 4.2, the GARCH model is used to predict a level of permanent income and this is used as the income measure (i.e. pt instead of yt). Again there is no major impact on the patterns. In Schmollers Jahrbuch 122 (2002) 3 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/schm.122.3.327 | Generated on 2023-04-04 12:30:24 Income Risk and Social Spending: Empirical Estimates 345 Table 6 Variations in regression structure Dependent Variable: Social Spending Independent Variables 1. Gini endogenous and instrumented 2. No fixed effects 3. No fixed effects, no instruments, no endogeneity beta s.e. beta s.e. beta s.e. Income .304 *.157 .359 M89 .265 *.077 Risk .571 .501 -.751 .851 -.066 .348 Inflation -.018 .038 .016 .050 .006 .035 Unemployment .386 *.109 .303 •.122 .292 *.051 Aged .936 *.282 .718 *.315 .927 *.074 Trade .042 .035 .101 *.035 .084 *.007 Union .020 .038 -.129 *.047 -.131 *.012 Gini -.124 *.047 -.006 .059 .024 .022 Strikes -.618 .819 -.596 1.312 -.626 .611 Mortality .126 .107 -.008 .117 -.022 .044 Turnout -.056 .039 .055 .052 .065 *.016 Military .420 .499 -.544 .358 -.414 *.117 Left -.050 .030 .142 *.041 .087 *.019 R2 0.7477 0.7388 0.7123 N 320 378 461 Addendum: Coefficient on risk in income equation .170 .283 .063 .556 -.424 *.195 Coefficient on social spending in income equation .147 *.076 .127 *.071 .061 *.022 Coefficient on social spending in risk equation .037 .022 -.019 .015 -.013 *.006 Note: coefficients identified with a are statistically significant at the 90 percent level, twotailed test. Source: Comparative Welfare States data set; Deininger and Squire inequality data. regression 4.3, log(yt) replaces yt as the income variable, again without major effects. In results not shown, regressions were run with various measures of income growth as the dependent variable in the income equation, again without any significant impact on the basic patterns. The results do not seem sensitive to the treatment of income. Table 5 shows some variations on the definition of risk and social spending. Regression 5.1 abandons the GARCH model and estimates risk simply Schmollers Jahrbuch 122 (2002) 3 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/schm.122.3.327 | Generated on 2023-04-04 12:30:24 346 Edward Castronova as squared deviations of log income around its time trend. This preserves the positive and statistically insignificant impact of risk on income in the income equation, but has little impact on the social spending equation. Regression 5.2 expresses social spending as a per capita figure, in thousands of real $US. Risk and strikes have a statistically significant and positive impact on spending under this definition, while the other effects are the same. Regression 5.3 expresses social spending as the share of non-health social spending in GDP. In this definition, strikes again have a positive impact and it is very large. Oddly, focusing on non-health social spending also makes infant mortality a positive motive for the Welfare State. Table 6 presents the results of more radical changes in the regression structure. First, one might argue that the level of inequality should be treated as endogenous. Also, the gini used here is derived from multiple studies using many methods, and is probably distorted by measurement error (although this would only affect the size, not the sign). Both problems require that the gini be instrumented. Regression 6.1 shows, however, that instrumenting the gini variable has no significant effect on its sign or magnitude, and it remains statistically significant. The negative impact of inequality on social spending seems to be both robust and causal in these data. Regression 6.2 explores the impact of ignoring the presence of fixed effects, and regression 6.3 also ignores the endogeneity of any variables (except the dependent variable) and abandons instrumenting. Here risk actually has a negative impact on social spending, although the effect is not statistically significant. Openness of the economy (Trade) seems to increase social spending while unionization decreases it. Interestingly, left voting here does increase social spending, which suggests that the common assumption that left parties support the Welfare State has its basis in the historical record prior to 1960 (i.e. the pattern of historical cross-country norms). The results in other tables suggest that this historical pattern is no longer valid. Other than this, the results are largely the same as in the other regressions. Considering all the regressions as a whole (and others not shown), the most robust findings are: - Social spending rises with income - Social spending rises with size of the aged and unemployed populations - Social spending falls with level of inequality Weaker results include: - Social spending increases risk taking - Risk increases social spending Schmollers Jahrbuch 122 (2002) 3 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/schm.122.3.327 | Generated on 2023-04-04 12:30:24 Income Risk and Social Spending: Empirical Estimates 347 - Left voting, strikes, military spending, turnout, and infant mortality have little apparent effect - There is no consistent effect of risk-taking on the income level, and only a small, statistically insignificant effect of social spending on risk-taking. 6. Conclusion The most robust results thus support the idea that demographics, more than anything, drive social spending. Sinn's insurance motive receives some support but it is only weak, while inequality, per se, does not have a strong influence. Future research could focus on two aspects of these results. First, the fact that social spending seems unrelated to poverty and inequality is itself surprising. Certainly, better data and better theories are needed to explain this counterintuitive, yet robust, outcome. Second, it is surprising that the results offer only weak support for the insurance motive. That may lie at the hands of the risk estimates, which seem to be noisy. Still, many different approaches were taken to measuring the variance of the income process, and none yielded tight estimates or robust regression coefficients. It would be ideal to develop a country panel data set with individual-level risk estimates. From a broader perspective, we might need to rethink what people view as "income risk." As researchers we tend to focus on risk as an observable component of the income process, the second moment of income. It is not clear that average people view their risks in such a manner, however, and it is their perceptions, and not our estimates, which affect behavior. To what extent does the second moment of income in a well-specified rational agent model of income determination accurately reflect the perceptions of income variability among real people? Since all estimates of income variance begin with the problematic expected utility model, we might not be surprised to find that our estimates of individually-perceived risk are unrealistic. These issues are similar to those that confront policy analysts attempting to design policies for handling environmental and workplace risk. It is not clear where risk perceptions come from, but they do not seem to come from a rational-actor expected utility model. The implication here is that large populations of aged and unemployed people, instead of the variance of income, might be the effective indicator of perceived income risk in the population. Whatever the true variance of his income, the citizen sees bread lines and imagines himself in them, and then votes for increases in social spending. This may or may not be a rational way to estimate the risk of poverty, but it Schmollers Jahrbuch 122 (2002) 3 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/schm.122.3.327 | Generated on 2023-04-04 12:30:24 348 Edward Castronova may be the way real people do it. If so (and this would be a good avenue for more work), the results here could be said to support the insurance theory as much as any other. References Aghion, Phillipe/ Caroli, Eve / Garcia-Penalosa, Cecilia (1999), "Inequality and Economic Growth: The Perspective of the New Growth Theories," Journal of Economic Literature 37,1615-60. Atkinson, Anthony B. (1995), Incomes and the Welfare State, Cambridge, Cambridge UP Atkinson, Anthony B. / Brandolini, Andrea (2001), "Promise and Pitfalls in the Use of 'Secondary' Data-Sets: Income Inequality in the OECD Countries as a Case Study," Journal of Economic Literature, 39(3), pp. 771-799. Baldwin, Peter (1990), The Politics of Social Solidarity and the Bourgeois Basis of the European Welfare State, 1875-1975, Cambridge: Cambridge University Press. Barr, Nicholas (1992), "Economic Theory and the Welfare State: A Survey and Interpretation," Journal of Economic Literature, 30(2), pp. 741-803. Becker, Gary S. (1985), "Public Policies, Pressure Groups, and Dead Weight Costs," Journal of Public Economics 28(3), 329-48. Bird, Edward J. (1995), "An Exploratory Comparison of Income Risk in the United States and Germany," Review of Income and Wealth 41(4), pp. 1-22. - (1999), "Politics, Altruism, and the Definition of Poverty," Journal of Comparative Policy Analysis 1(3), pp. 269-91. - (2000), "Does the Welfare State Induce Risk-Taking?", Journal of Public Economics, June 2001. Cameron, David (1978), "The Expansion of the Public Economy: A Comparative Analysis," American Political Science Review, 72,1243 - 61. Deininger, Klaus / Squire, Lyn (1996), "A New Data Set Measuring Income Inequality," World Bank Economic Review 10(3), 565-91. Esping-Anderson, Gosta (1990), The Three Worlds of Welfare Capitalism, London: Polity Flora, Peter /Heidenheimer, Arnold J. eds. (1981), The Development of Welfare States in Europe and America, London: Transaction. Greene, William H. (1993), Econometric Analysis 2e, New York: Macmillan. Hicks, Alexander M./Swank, Duane H. (1992), "Politics, Institutions and Welfare Spending in Industrialized Democracies, 1960-82," American Political Science Review 86(3), 658-74. Hicks, Alexander M. / Misra, Joya (1993), "Political Resources and the Growth of Welfare in Affluent Capitalist Democracies, 1960-1982," American Journal of Sociology 99(3), 668-710. Schmollers Jahrbuch 122 (2002) 3 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/schm.122.3.327 | Generated on 2023-04-04 12:30:24 Income Risk and Social Spending: Empirical Estimates 349 Himmelfarb, Gertrude (1992), Poverty and Compassion: The Moral Imagination of the Late Victorians, New York: Vintage. Hochman, Harold M./Rodgers, James D. (1969), "Pareto Optimal Redistribution American Economic Review 59(4), pp. 542-57. Huber, Evelyne/Ragin, Charles/ Stephens, John D. (1993), "Social Democracy, Christian Democracy, Constitutional Structure, and the Welfare State," American Journal of Sociology 99(3), 711-49. - (1997), Comparative Welfare States Data Set, Northwestern University and University of North Carolina. Katzenstein, Peter J. (1985), Small States in World Markets: Industrial Policy in Europe, Ithaca: Cornell University Press. Kristov, Lorenzo/Lindert, Peter H./McClelland, Robert (1992), "Pressure Groups and Redistribution," Journal of Public Economics 48, 135 - 63. he Grand, Julian (1987), Not Only the Poor: the Middle Classes and the Welfare State, London: Allen and Unwin. Lindert, Peter H. (1994), "The Rise of Social Spending, 1880-1930," Explorations in Economic History 31,1-37. - (1996), "What Limits Social Spending?" Explorations in Economic History 33, 1- 34. Mead, Lawrence M. (1997), The New Paternalism, Washington: Brookings. Meitzer, Allan H./Richard, Scott F. (1981), "A Rational Theory of the Size of Government," Journal of Political Economy 89(5), 914-27. Pampel, Fred C. and John B. Williamson (1989), Age, Class, Politics, and the Welfare State, Cambridge: Cambridge University Press. Persson, Torsten /Tabellini, Guido (1994), "Is Inequality Harmful for Growth?" American Economic Review 84(3), 600-21. Pierson, Paul (1996), Dismantling the Welfare State? Cambridge: Harvard University Press. Piven, Frances Fox/ Cloward, Richard A. (1971), Regulating the Poor: The Functions of Public Welfare, New York: Vintage. Rimlinger, Gaston (1971), Welfare Policy and Industrialization in Europe, America, and Russia, New York: Wiley. Sinn, Hans-Werner (1995), "A Theory of the Welfare State," Scandanavian Journal of Economics 97(4), 495-526. Temple, Jonathan (1999), "The New Growth Evidence," Journal of Economic Literature 37, 112-56. Trattner, Walter I. (1999), From Poor Law to Welfare State, 6e New York: Simon and Schuster. Uusitalo, Hannu (1984), "Comparative Research on the Determinants of the Welfare State: the State of the Art," European Journal of Political Research 12(4), 403 - 22. Visser, Jelle (1996), Unionisation Trends. The OECD Countries Union Membership File, Amsterdam: University of Amsterdam, Centre for Research of European Societies and Labour Relations CESAR. Schmollers Jahrbuch 122 (2002) 3 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/schm.122.3.327 | Generated on 2023-04-04 12:30:24