The Autocratic Welfare State: Resource distribution, credible commitments and regime survival

The Autocratic Welfare State: Resource distribution, credible commitments and regime survival November 22, 2015 Abstract We investigate the adoption, characteristics and consequences of social policies in autocracies. Our argument on why autocrats implement such policies, particularly pension programs, highlights that they can be considered club goods that I) are targeted to critical supporting groups and II) solve credible commitment problems on promises of future distribution, thereby mitigating the probability of regime breakdown. We test different implications from the argument by combining extant regime data with novel data on welfare state programs, for 140 countries with time series from the 1880s. We find that autocracies are as likely as democracies to have pension systems. But, autocracies have less universal pension (and other) programs. Further, we find that such programs reduce probabilities of autocratic breakdown and of democratization.

1

Introduction

The welfare state literature has considered social policy programs either as mechanisms for providing individuals with insurance against risks such as job-loss or illness, or as tools for progressive redistribution (see Moene and Wallerstein 2001). Hence, different theoretical models are well suited to explain the existence of social policy programs in democracies; most voters are risk averse (Chetty 2006), thus preferring even costly insurance, and the (relatively poor) median voter often has incentives to vote for parties promising income redistribution (Meltzer and Richard 1981). Yet, social policy programs are adopted not only in democracies, but also in autocracies. The perhaps most famous examples are the sickness, accident, disability and old-age pension programs adopted in Bismarck’s Germany, but these are not exceptions. Mares and Carnes (2009, 97) count that the vast majority of the 38 “non-industrialized” countries for which they have data actually adopted their first old-age pension system under autocracy. Indeed, analyzing more than 140 countries, with time series from the 1880s, we show below that autocracies are no less likely than democracies to adopt or have old-age pension systems. But, why would autocrats adopt such programs and spend resources on pensions rather than keeping tax revenues for private consumption, or for investment in repressive capacity to ensure regime survival? We propose that pension programs in autocracies are, indeed, used for political survival purposes, and our statistical tests estimate a clear and sizable causal effect of pension programs on regime survival once adjusting for the endogenous adoption of such programs. Social policies, and notably old-age pensions, involve targeted distribution of resources to particular groups, and thus constitute a special form of co-optation. While the role of formalized social policy programs has received comparatively little attention in the literature on co-optation in autocracies, which often focuses on discretionary distribution and patronage networks, this insight is not new. Mares and Carnes (2009), for instance, explicitly argue that social policies are used for targeted distribution to critical groups. However, neither Mares and Carnes nor the wider literature have elaborated on a second critical feature of such programs, namely how they allow autocrats to make credible promises of future distribution. We thus respond to the theoretical challenge raised by, for example, Acemoglu and Robinson (2006); why would any group trust a dictator’s current promise of future distribution (see also, e.g., Svolik 2012)? In brief, we conceptualize social policy programs in autocracies as club-goods, and highlight that

1

they combine two characteristics – i) inter-temporal credibility and ii) potential for targeting – which make them vital for autocratic survival. First, in contrast with discretionary distribution of private goods, social policy programs, such as old-age pension programs, are fairly transparent, stable and predictable, and not easy to reverse without substantial costs for rulers. Hence, distribution through these programs – not unlike establishing particular institutions such as ruling parties (Magaloni 2008; Gehlbach and Keefer 2011, 2012; Svolik 2012; Boix and Svolik 2013) – may solve the critical credible commitment problem. Second, in contrast with pure public goods, pension programs can be targeted towards identifiable groups. Hence, autocrats can design programs channeling resources to groups they need support from to stay in power, rather than spending precious resources on irrelevant groups. Although implementing and sustaining, for example, old-age pension programs could be more resource-demanding than discretionary distribution, it is thus preferable under some circumstances, for instance when the critical supporting groups of the regime are fairly large. We thereby also address the important question of why (and under what circumstances) dictators would opt to distribute through targeted social policy programs rather than targeted discretionary spending. The argument generates various empirical implications – including on the typical characteristics of autocratic pension programs, and how they affect regime survival. We test these implications using our new Social Policies Around the World (SPAW) dataset, with global coverage and time series from the 19th century. By detailing how and why autocrats employ pension programs, we bridge (and contribute to) the welfare state literature, which has mainly focused on (developed) democratic contexts, and the rapidly growing literature on autocratic politics. After having briefly reviewed relevant literature on social policies and survival strategies in autocracies (Section 2), we present our argument (Section 3). We thereafter describe our novel SPAW dataset (Section 4). In the empirical analysis (Section 5), we report evidence that autocracies are as likely as democracies to adopt and have pension systems, although autocracies have less universal systems. Finally, we present the first systematic empirical study of its kind on how social policies affect regime survival and change, finding clear evidence that pension systems reduce the probability of autocratic regime breakdown. This estimated effect, which holds in models trying to account for the implementation of pensions being endogenous to the security threats facing the regime, is statistically significant and substantial in size. In extension, we report evidence suggesting that implementing pension systems may not only make autocracies less likely to break down, in general, but also reduce

2

the probability of democratic transitions.

2

Literature

Most welfare state researchers have focused on OECD democracies (Flora and Heidenheimer 1981; Korpi 1983; Esping-Andersen 1990; Huber and Stephens 2001; Iversen 2005). Different theoretical accounts then also suggest that social policies – and the provision of public services such as basic health-care and primary education (e.g., Lake and Baum 2001; Stasavage 2005; Harding and Stasavage 2014) – should be more widespread, better financed, and have broader coverage in democracies than in autocracies (Boix 2003; Acemoglu and Robinson 2006; Haggard and Kaufman 2008; Ansell 2010). There is evidence that Western countries expanded welfare spending due to franchise extensions from the 19th century onwards (Lindert 2005; Kim 2007). Yet, some empirical studies report that democracies are not, on net, associated with higher social spending (e.g., Mulligan, Sala-i Martin and Gill 2003).1 Former Communist dictatorships presided over extensive social policy programs (Milanovic 1998), but so did autocracies such as PRI-Mexico (Magaloni 2006). Desai, Olosg˚ard and Yousef (2009) find that – among autocracies – those with harsher repression provide more welfare spending. Famously, a historical pioneer in implementing welfare policy programs was Imperial Germany, under the strategic guidance of Chancellor Bismarck (see Rimlinger 1971). Different strands of literature provide two particularly important insights about autocratic politics. We combine these two insights in our argument on how autocrats conduct distributive politics, and apply them more specifically to old-age pension systems: First, policies are intentionally used by many autocrats to remain in power. Second, ensuring that promises made to the regime’s critical supporters are credible is of utmost importance: Regarding the first insight, some autocrats might be genuinely concerned with the welfare of their citizens (see, e.g., Wintrobe 1998, 95–96), and provide old-age public pensions or unemployment insurance to improve their life quality. However, the literature often considers policy-making as a strategic tool used mainly for for enhancing regime survival (e.g., Bueno de Mesquita et al. 2003; Svolik 2012). Several studies show how autocratic regimes rely on particular groups for their political survival, and design economic policies to satisfy these groups. Bueno de Mesquita et al. (2003) theorize that whereas large-coalition rulers should find it cost-effective to spend resources on 1

There are likely systematic differences among democracies; proportional representation (PR)

electoral systems may enhance social spending and induce more universal programs than pluralmajoritarian (e.g., Persson and Tabellini 2003; Iversen and Soskice 2009). 3

public goods, small-coalition rulers should allocate private goods directly to their few vital supporters. Democracies, because of larger coalitions, should observe more spending on public goods than dictatorships (see also Lake and Baum 2001). Still, there are differences in coalition size between dictatorships – for example, party-based autocracies generally have larger coalitions (Fjelde 2010) – and thus also in optimal resource-distribution strategies. Also regarding social policies, more specifically, regime survival was widely considered a core motivation for Bismark’s reforms (Rimlinger 1971), and Mares and Carnes (2009) and Haggard and Kaufman (2008) outline how (targeted) social policies were designed and implemented, at least in part, for political-survival reasons in various autocracies. Regarding the second insight, the theoretical argument in Acemoglu and Robinson (2006) highlights how autocratic elites often cannot credibly commit to future redistribution, even when faced with a revolutionary threat; they would expectedly stop redistributing (i.e., stop “wasting resources”) once the threat is removed. The dictator may ensure survival through repression, but this is costly (Wintrobe 1998), and co-optation is often preferable (see Gandhi 2008). Although autocrats may face difficulties in making promises of future distribution credible, this is far from impossible. Different studies (North and Weingast 1989; Geddes 1999; Magaloni 2006; Gandhi 2008; Wright 2008; Gehlbach and Keefer 2011, 2012; Svolik 2012; Boix and Svolik 2013, for democracies, see Iversen 2005) point out that autocrats can reduce credible commitment problems through establishing regime parties and other institutions, such as legislatures. By allowing, for example, party institutions to check the ruler, the latter cannot easily alter distributive policies favoring the current supporting coalition. Hence, promises of future distribution are more credible within party-based regimes, in turn reducing risks of regime splits, non-constitutional leader removal, or even regime collapse (Geddes 1999). However, we take this one step further by proposing that not only institutions, but also the design of distributive policies matter; formalized social policy programs – and particularly old-age pension programs – mitigate credible commitment problems.

3

Why would autocrats implement social policies?

We argue that autocrats often have strong incentives to implement social policy programs to please their supporters, and this effectively helps them prolong their rule. Institutionalizing the distribution of resources through such programs makes future distribution to particular groups more credible than promises of discretionary private-goods distribution. When distributive policy is clearly codified, the

4

autocrat cannot as easily backtrack on promises once the relevant group no longer poses an imminent threat. Changing social policy programs may be particularly difficult in institutionalized autocracies with considerable checks on the ruler, as reversing programs require the co-operation of multiple veto players. Implementing generous welfare programs does imply unwelcome future costs. Yet, this is something the autocrat willingly accepts if the alternative is losing power; when promises of future distribution are considered credible, the benefiting groups will more likely support the regime. Program benefits should, however, be targeted towards groups the regime needs support from – be it urban workers or soldiers – whereas “irrelevant” citizens should receive little. By appeasing groups that might otherwise contest the regime, the introduction of such programs should bolster autocratic regime survival, and even worsen prospects for democratization. Old-age pension systems should be particularly suitable as such instruments of regime survival. First, the amount of resources channeled through pension programs often dwarfs the amount channeled through other programs, such as unemployment benefit programs. Second, old-age pension systems can easily be designed to target various groups. Pension schemes are more flexible cooptation devices than, say, maternity leave programs; pensions are not restricted to pre-given groups, such as women of child-bearing age, that may or may not be relevant for political survival. Also, unemployment benefits programs advantage those with high risks of losing their jobs, and autocracies have typically focused on ensuring the job safety of critical supporting groups through employment protection legislation and other means (e.g., Mares and Carnes 2009). Further, disability insurance might come in handy for members of the autocrat’s critical supporting groups, but the risk of disabilitating work injury is relatively small in comparison with that of growing old. Third, old-age pension systems speak directly to the temporal dimension of the credible commitment problem. For most concerned individuals, pensions will be paid out in the not-too-near future – i.e., when current coalition members are retired from their positions, and thus perhaps less capable of holding the autocrat accountable. Absent old-age pension schemes, current coalition members may not trust dictators to provide future resources. Let us explore the argument and its core assumptions in more detail. To simplify, we divide the population in dictatorships into two groups, the regime’s critical supporters (S) and all other citizens (C). In addition, we have the regime (R), i.e. the dictator and his/her very close collaborators. S may, for instance, consist of soldiers, land-owners, civil servants, former army veterans, or workers in vital

5

industrial sectors. One key aspect, however, is that members of S are identifiable according to some feature (e.g. “Participated in war of independence” or “Works in the oil industry”). Leaving repression aside, we consider co-optation through current and (promised) future distribution of resources the key policy tool R employs to stay in power.2 Since, R needs the approval of S, but not C, to remain in power, R is concerned only with pleasing the former. Members of S care about present and future consumption, and thus require resources distributed to them today and credible promises of resources tomorrow. If not, S may seek replacing R with another regime more willing to comply with their demands, or even pursue democratization which has some benefits also to economic and other elites (Ansell and Samuels 2014). R can basically co-opt S through three strategies: 1) lump-sum provision of private goods; 2) public goods provision benefiting both S and C; and, 3) provision of club goods to S. Club goods are non-rivalrous; their consumption have externalities, and club goods thus affect a group of individuals. But, in contrast to public goods, they are excludable. Pension checks are posted to individuals and consumed by them privately – distribution through pensions therefore has some private-good characteristics – but there are good reasons to consider segmented pension programs as club goods: 3.1

Social policies as club goods – targeting

One key aspect of social policies is their insurance function (Moene and Wallerstein 2001). Individuals cannot perfectly predict whether they will be unemployed five years from now or capable of working when they turn 70, and many would sacrifice some current income to insure against the low income levels accompanying such events. Payments should, importantly, be implemented under clearly specified conditions and in a rule-following manner. If the government could simply decide on a case-by-case basis which members should (not) receive payments, this would undermine the program’s general credibility and reduce expected utility for all program members. Hence, the insurance aspects of welfare programs have non-rivalrous properties. Regarding excludability, this requires the presence of observable, codifiable criteria for distinguishing program members from non-members. The excludability aspect of pension programs, combined with R’s incentives, indicates that such programs will be targeted (only) towards members of S (see also Mares and Carnes 2009). Haggard and Kaufman (2008) describe the targeted character of welfare states in numerous Asian and Latin American countries. In Thailand and South Korea, for 2

Yet, repressive strategies are likely endogenous to autocrats’ co-optation strategies (Desai,

Olosg˚ard and Yousef 2009; Frantz and Kendall-Taylor 2014). 6

example, social insurance was limited “to core constituents in the state itself: the military and civil servants” (p. 140), whereas in pre-Per´on Argentina “social-insurance protection was confined mainly to the military, white-collar workers, and a few strategically situated blue-collar unions” (p. 95). R has strong incentives to under-provide truly public goods, since these are only cost-effective in ensuring political survival when the critical supporting groups are very large (Bueno de Mesquita et al. 2003). This leaves R with the two targeted co-optation strategies, discretionary distributing private-goods to individuals in S or establishing club goods – notably targeted pension programs – for S collectively. The former has some important benefits from R’s point of view. First, if R, at some point, strongly consolidates power, R can easily reduce the amount of private goods distributed, leaving more resources for R’s personal consumption. Formalized social policy programs are, as discussed below, more difficult to adjust at will. Thus, dictators (and their closest collaborators) who consider themselves very safe in power should prioritize private-goods distribution. Second, discretionary distribution of private goods is arguably easier to direct exclusively to members of S; in practice, social policy programs may lead to “leakage” or “waste”; some citizens that are unimportant for the dictator’s survival may receive benefits. For instance, former servicemen with little political influence are covered by veteran benefits, or “irrelevant” parts of the manufacturing sector producing in the periphery are covered by pension schemes for manufacturing workers, even if these are mainly intended for powerful firms and workers in the capital. 3.2

Social policies and solving commitment problems

Yet, R, particularly when the future grip on power is insecure, may substitute private goods distribution with formalized and fairly stable social policies targeting critical groups. The reason is the exact same stability of such programs making them monetarily more costly. R wants to avoid making S uncertain about future distribution; such uncertainty may induce S to replace R. Acemoglu and Robinson (2006), among others, highlight the fundamental problems of time consistency and credibility of promises facing autocrats. They propose that a liberalizing regime change, thus giving away political decision-making powers, is the only viable solution for autocrats facing imminent threats. Still, recent contributions have highlighted how particular institutions in autocracies alleviate this problem (Svolik 2012; Boix and Svolik 2013). We further propose that the way in which distributive policies are designed contribute to determine whether promises of future distribution are credible or not. If the future reversal of policies entail high economic or political costs, autocrats can credibly

7

promise their continuation. Thus, autocrats may want to “tie their own hands” when designing policies that channel resources to critical groups. One way of doing this is through codifying pension programs. Implementing such programs stabilizes distributive policy through multiple mechanisms. For instance, R initially makes investments in administrative and monitoring capacity, thus reducing (the relative) costs of continuing the program once up and running, creating lock-in effects (Pierson 2000). Further, the reversal of pension programs constitute easily identifiable events which could function as signals, or “sparks” (Kuran 1989), alleviating anti-regime collective action problems. The relative ease with which members of S can identify “breaches to co-operation” reduces R’s incentives to reverse such programs to begin with. Thus, pension programs, once implemented, constitute “guarantees” of future distribution, thereby reducing coalition incentives to remove the regime. Indicatively, Ponticelli and Voth (2011), using European data from 1919–2008, find that cuts in public spending generate unrest and anti-regime collective action. Our SPAW data also show that social policy programs are durable and seldom retracted or scaled down substantially. For instance, only 10% of changes to old-age pension systems in our dataset, in terms of expanding or reducing the number of groups covered, are reductions. Also following expectations, such reductions in pension programs are far less likely in established autocracies than immediately after regime change when autocrats may cut programs benefitting supporters of the old regime (see also Albertus and Menaldo 2012) – the relative frequency of such pension reductions is almost three times higher in the first three years after regime change. Withdrawing pension benefits from S is likely very dangerous for R. Anticipating this, members of S perceive promises of future resource distribution as credible if linked to pension programs. This is the case particularly if members of S expect the regime to be long-lived – which is the case especially for party-based autocracies (e.g., Geddes 1999). Hence, S will also have incentives to support R’s hold on power once a pension program is in place. 3.3

Empirical implications

The argument produces quite diverse implications, and we highlight three hypotheses. First, since costs of setting up and monitoring policy programs are likely increasing but concave in number of recipients, small-coalition dictatorships may more often employ non-formalized distribution of private goods to their supporters, and dictatorships with larger winning coalitions should, everything

8

else equal, more often operate social programs than small-coalition dictatorships.3 However, our argument does not imply that the systems with the largest coalitions, namely democracies, have more programs. Although large coalitions increase incentives to spend through formalized programs, this could very well be balanced by democratic leaders not needing to solve the theoretically highlighted commitment problems through such mechanisms. Leaders and voters know the former may more easily be removed from office in the next election if they do not redistribute (e.g., Acemoglu and Robinson 2006). On the other hand, such commitment issues were at the core of our argument for why autocracies operate welfare programs. More specifically, our expectations on (lack of) clear and systematic differences between democracies and autocracies mainly concern old-age pension systems. Pension schemes are, as discussed, more flexible autocratic co-optation devices than, for example, maternity leave programs, and for most concerned individuals pensions will be paid out only after several years – i.e., when current coalition members are retired from their positions, and thus less capable of holding the autocrat accountable. Absent old-age pension schemes, current “coalition members” may not trust dictators to provide future resources. Thus, the coalition size and commitment mechanisms may very well balance each other out, in terms of inducing leaders to provide pension systems in autocracies and democracies. One plausible hypothesis is therefore: H1) Autocracies are not (systematically) less likely than are democracies to adopt/have old-age pension systems. Yet, our argument clearly indicated that the social policies, notably including pension programs, pursued in autocracies should, generally, be more targeted than in democracies. This mainly stems from autocracies having narrower supporting coalitions. This yields H2) Old-age pension systems are less universal in autocracies than in democracies. In extension we also test whether, party-based autocracies, which often have larger coalitions (but which are also associated with other factors that may affect welfare characteristics in different 3

Extant measures only allow us to test this indirectly: Party-based autocracies typically have larger

coalitions (see, e.g., Fjelde 2010), and one might therefore expect them to more often operate welfare programs. We test this empirically below. However, one theoretically relevant moment that may attenuate this net association is that some ruling party institutions could be so strong that they, in themselves, ensure promises of future distribution being credible (e.g., Svolik 2012), reducing the need to formalize distributive policies through pension programs.

9

directions), have more universal programs than monarchies or military regimes. Factors influencing the risk of regime breakdown might incentivize autocrats to initiate (or expand) pension programs. However, forward-looking autocrats may implement such programs as safety-nets long before the economy crashes or protesters assemble in the streets (when it may be too late to save the situation by promising benefits). Nevertheless, the notion that such programs – once in place – boost autocratic survival prospects is a testable proposition following our argument. Since we highlight that adoption of pension programs is endogenous we must adjust for selection mechanisms when estimating their effects on survival. Hence, our final hypothesis is H3) Old-age pension systems reduce the risk of autocratic regime breakdown Further, if social policy programs reduce commitment problems for autocrats, as we argue, they should expectedly also mitigate the (more specific) probability of democratizing regime transitions (see Acemoglu and Robinson 2006). As another extension we thus estimate the effect of old-age pension systems on democratization.4 Also, despite our focus on old-age pensions, we elaborate on various nuances of the “Autocratic Welfare State” by briefly discussing resembling tests on the five other types of social policy programs coded in SPAW.

4

The SPAW dataset

Our new Social Policies around the World (SPAW) dataset covers 223 countries, with time series for some variables from 1790–2010. SPAW draws on various sources, notably the ILO Legislative series (1919-) and US Labor Department SSPTW-reports (1937-).5 As detailed below, it contains variables measuring first year of legislation and existence of programs and which groups are covered for all major transfer programs, allowing us to test whether autocrats implement particular social policies, such as targeted pension programs, and whether these, in turn, systematically affect regime survival. Previous analysis of determinants or effects of social policies largely rely on aggregated spending measures. While interesting for some purposes, they are often poor proxies for the welfare-state characteristics of proposed causal relevance (see, e.g., Esping-Andersen 1990; Scruggs 2006). We are here interested in capturing to what degree benefits are targeted to specific groups, or welfare state segmentation. Since spending measures tell very little about the structure of benefits, they are 4

This empirical implication contrasts directly with, for instance, that following from the argument

in Ansell and Samuels (2014), proposing that targeted social welfare spending should enhance the probability of democratization (see p.144). 5 For details, see the Appendix and the SPAW codebook at REMOVED WEBPAGE/REFERENCE.

10

unfitting for this purpose; two countries can have identical spending levels but very different welfarestate structures (Esping-Andersen 1990; Allan and Scruggs 2004). Likewise, an important alternative measure – the population share insured – is also problematic. For example, insurance statistics will not distinguish countries where only one small group, say urban industrial workers, is insured, from another where benefits accrue to all workers but the required contribution period is so high that only some end up eligible. In the first case benefits are targeted to a specific group, whereas in the latter eligibility is more “random”. Consequently, neither spending nor insurance measures are very appropriate for analyzing how autocrats use targeted benefits to survive by reducing regime threats from particular groups, such as urban workers. Different variables in SPAW alleviate such problems. The variables are grouped into two main categories, I) legislative changes and II) program characteristics. For both I) and II), SPAW contains information on sickness leave, maternity leave, unemployment, disability/work injury, family allowances, and old-age pensions. For legislative-changes, we identify first year of major reform and presence of (above-minimumthreshold) programs. Contrary to previous attempts to capture first year of legislation (e.g., Mares and Carnes 2009), we distinguish between major and special programs (see Appendix for closer discussion). Several countries legislated special public servant pensions during the 19th century, way before introducing old-age pension programs in their modern form. In Argentina, for instance, the first civil servant pension (originally just for the judiciary and treasury) was enacted in 1803, whereas the first law covering industrial workers passed 143 years later in 1946 (Mesa-Lago 1978). Our operationalization thus only considers a (national-level) pension system to exist when at least one major social/occupational group (see below) is covered. The variables in SPAW improve on previous codings in several regards. First, much effort has gone into using multiple sources – both cross-country reports and country-specific sources – to crosscheck the validity and reliability of the codings. While this introduces and element of subjective judgement even for such de jure measures, this is highly preferable to employing only one of the sources at hand, both for increasing coverage and for avoiding different issues associated with the individual sources. Second, the main source is the ILO Legislative series, which we deem as the most reliable. Previous attempts to measure “first laws” have tended to use SSPTW reports. The latter are associated with different problems, for instance operating without a consistent criterion over time (or between countries) on what constitutes a first law. Depending on the edition, the SSPTW reports

11

switch between using special programs or major programs. Thus, we have employed the entire archive of SSPTW reports, alongside the other sources, and not only a restricted sample. Regarding program characteristics, these may be conceptualized as varying along two dimensions, IIa) eligibility and IIb) distributive potential. Eligibility refers to rules defining who can partake, whereas distributive potential concerns the program’s inherent capacity to allocate benefits to members. Therefore, eligibility relates to who is covered against risks, and distributive potential to the benefits program members can expect. SPAW measures eligibility by first classifying whether programs are means-tested, employment-based or fully universal, and make further nuances by constructing Universalism indices (see below) – with the longest time series starting with the German accident liability insurance of 1871. For distributive potential, SPAW contains numerous variables, including measures of whether program eligibility is constrained by income ceilings; number of days recipients must wait before receiving benefits; number of weeks program benefits extend over; and, whether the state subsidizes programs. We use two types of variables from SPAW below, measuring particularly relevant aspects for our argument. The first are dichotomous measures of whether pension (and other social policy) programs – passing a minimum threshold – exist or not. This minimum threshold is operationalized as the program covering ≥ 1 of the following 8 major social/occupational groups: agricultural workers; industrial/production workers; small-firm workers; self-employed; students; employers; temporary/casual workers; family/domestic workers.6 The second are Universalim Indices, capturing what parts of the population are included on equal terms in the same program.7 The baseline versions range from 0–9, and are scored 0 if programs are absent and 1 if programs are means-tested based on some property criteria (income-based exclusions are not counted as means-tested programs; as robustness tests we run models excluding means-tested 6

Even laws (generally) prescribing broad coverage might simultaneously exclude specified groups

such as ethnic minorities (or majorities, as in Apartheid South Africa) or certain occupational groups. SPAW thus includes a measure capturing number of explicitly excluded groups. Second, although covered in general systems, groups might receive special benefits from additional programs. In SPAW, the (country-year) average is 3 special programs, with a maximum of 63 special programs (Greece 1983). 7 Below, we do not distinguish voluntary from compulsory coverage, or social insurance from private schemes (provident funds), but SPAW includes variables coding, e.g., whether or not systems are private.

12

programs and our results remain almost identical). Further, contribution- or employment-based programs covering one major group are scored 2. 3-scores indicate that programs are contribution- or employment-based and cover two major groups, and so on, up to 8-scores where seven groups are covered. Maximum 9-scores indicate all residents are automatically entitled to benefits – a fully universal system.8 Appendix Figure A.1 shows the baseline Universalism Index distributions for the six major programs. Although this index is not of interval scale, we employ easy-to-interpret linear regression models for our baselines. We also tested ordered logit and probit models, and logit and probit models run on dummy variables (constructed by using different cut-offs) and our results are robust (see Appendix). While the SPAW data allow us to better measure theoretically important characteristics, there are potential reliability and validity issues also with these data. First, while coding de jure programs allows for constructing an extensive database with comparable measures across time and space, some programs may be “empty shells”. Hence, our data could, for instance, include pension programs that exist on paper only, but do not distribute resources in practice. While we would have loved to conduct systematic tests on de facto program characteristics, these are very difficult to code in a valid manner. Further, if many programs are simply empty shells, they would expectedly not fulfill the abovetheorized roles and have a clear causal effect on regime survival. Hence, our strong results below may actually underestimate the “true effect” of pension programs on regime survival. Likewise, de facto dismantling of existing programs (see Mares and Carnes 2009, 94) should attenuate our estimates. Second, we do not differentiate groups according to size. In principle, one could weight scores by group size, but there is no available, sufficiently disaggregated information for coding groups in a comparable manner between countries or over time. We therefore use non-weighted measures. In any case, our theoretical argument evolves around whether benefits are targeted to few or many (clearly identifiable) groups. Third, our Universalism Indices do not make explicit qualitative distinctions between the particular groups covered. Expectedly, some major groups – such as domestic workers – are constituencies autocrats are unlikely to grant benefits, whereas other – such as industrial workers – are critical supporting groups in many contexts. In democracies, leaders may need to cater to both of these groups to retain power, and autocracies should thus have less universal systems than democracies (while both should expectedly have pension programs covering at least one group). Our results below – indeed indicating higher universalism scores among democracies – would not reflect 8

If, e.g., different pension programs exist, all are coded and we average over these programs. 13

our argument well if pension systems were first enacted covering only domestic or family workers (while excluding, e.g., industrial workers). This is, however, not the case in practice. During our coding, we found no instances of coverage being extended to domestic/family workers without urban manufacturing/industrial workers also receiving benefits.9 In sum, the measures used capture our theoretical concepts quite well.

5

Empirical analysis We organize our tests in three sub-sections. First, we test H1 on regimes and adoption of pension

systems. We then test H2 on how targeted or universal pension programs are under different regimes. Finally, we test H3 by investigating how pension programs causally affect autocratic regime breakdown. Corroborating our expectations, we, for instance, find that autocracies are no less likely to have pension programs than democracies, but that democracies generally have more universal such programs. To the the extent that our models allow for causal identification (which is, of course, always difficult to ensure in a cross-country time-series framework), we find strong evidence of the expected effect of pension programs on autocratic survival, but also on democratization more in particular. 5.1

Does regime type matter for probability of adopting pension programs?

We argued that autocrats use social policies for targeting and credibly promising resources to groups that are critical to regime survival. In particular, old-age pension schemes can be effectively targeted towards groups capable of affecting regime survival. Hence, while democracies and autocracies may adopt pension programs for different reasons, we do not expect any clear and systematic differences in the probabilities of adopting or having such programs between regime types. To test H1, we mainly use the dichotomous Boix, Miller and Rosato (2013) regime measure (henceforth BMR). BMR distinguishes between democracies and autocracies according to existence of free and fair elections and whether more than half of the male population can vote. The dependent variable in Models A1–A3, Table 1, is a dummy recording whether an old-age pension system (covering ≥ 1 major occupational/social group) exists (whereas Models A4–A6 have adoption of pension system as dependent variable). BMR and existence of pension system correlate by .26 (9166 obs. from Model A1); democracies more often have pension systems than autocracies, although the 9

Benefits were usually first extended to urban employees such as industrial, commercial workers,

or salaried employees, with or without restrictions on firm size. Where agriculture was also included, a distinction was often made between workers using and not using electric power, ensuring only workers close to urban centers were covered. 14

difference has dwindled in more recent decades. This is illustrated by the upper-left graphic in Figure 1, which shows shares of democracies and autocracies with pension programs.

.8 .4 .2

0

0

.2

.4

.6

.8 .6

.8 .6 .4 .2 0

Sickness Benefits 1

Maternity Benefits 1

1

Old-age Benefits

1900 1920 1940 1960 1980 2000

Accident Benefits

Unemployment Benefits

Familiy Benefits

1900 1920 1940 1960 1980 2000

0

.2

.4

.6

.8

1 0

0

.2

.2

.4

.4

.6

.6

.8

.8

1

1900 1920 1940 1960 1980 2000

1

1900 1920 1940 1960 1980 2000

1900 1920 1940 1960 1980 2000

Democracies

1900 1920 1940 1960 1980 2000

Autocracies

Figure 1: Shares of democracies and autocracies, as operationalized by Boix, Miller and Rosato (2013), with programs in six social policy areas during the 20th century. Nevertheless, this correlation ignores other factors affecting adoption of pension systems. Pension systems expanded during the 20th century, although this occurred earlier in some regions. Since also democratization has clustered in temporal waves, differently timed across regions (see Huntington 1991), controlling for time period and geographic neighborhood is important. Our models therefore always include year dummies and eight region dummies (from Miller 2013). We also include log GDP per capita (from Maddison 2007); richer countries have more resources to spend on pension systems, and income correlates with democracy. For our baselines, we run logit models on existence or adoption of pension system, with BMR as the key independent variable, and with errors clustered on country to account for panel-specific autocorrelation (which is also alleviated by the year dummies). However, we also tested Cox survival models and logit models using a dummy variable BCSTS specification following Beck, Katz and Tucker (1998) or cubic splines following Carter and Signorino (2010) to further account for temporal dependence. Our baseline logit results are fairly robust to these and other plausible changes to the model set-up (see Appendix).10 10

Ideally, we would control for country-specific effects in our logit regressions, but this is infeasible.

We include country dummies below when investigating degree of universalism, which provides far more within-country variation. 15

Democracy (BMR) Ln GDP p.c. Ln population Ethnic fractionalization Urbanization Military size Resource dependence Region dummies Year dummies N Countries

A1 A2 A3 DV: Existence pension program b/(t) b/(t) b/(t) 0.128 0.095 -0.010 (0.32) (0.24) (-0.03) 0.498* 0.693*** -0.123 (1.75) (2.59) (-0.42) 0.445*** 0.424** (2.85) (2.41) -0.615 -1.040 (-0.62) (-1.09) 0.063*** (3.94) -0.010 (-0.05) 0.001 (0.06) Y Y Y Y Y Y 9166 8330 7881 140 137 137

A4 A5 A6 DV: Adoption pension program b/(t) b/(t) b/(t) 0.353 0.437 0.350 (1.11) (1.32) (1.09) 0.098 0.190 -0.178 (0.62) (1.03) (-0.59) 0.305** 0.317** (2.55) (2.47) -0.273 -0.842 (-0.46) (-1.27) 0.029*** (2.70) 0.045 (0.51) 0.004 (0.25) Y Y Y Y Y Y 1504 1344 1252 114 111 107

Table 1: Democracies, autocracies, and existence or adoption of pension systems. Notes: *** p < 0.01, ** p < 0.05, * p < 0.10. Logit regressions with errors clustered on country. Existence of minimum old-age pension system is dependent variable in A1–A3 and adoption in A4–A6 (here, countries already having systems are omitted). Year dummies, region dummies and constant are omitted from table. Maximum time series extend from late 1880s–2004.

Model A1, using these controls, is our most parsimonious specification. When accounting for income and region- and year-specific effects, the correlation between democracy and pension system vanishes. BMR is positive in A1, but t = 0.3. Hence, we find initial support for H1 – there is no systematic difference between democracies and autocracies regarding existence of pension systems. Model A2 adds two potentially relevant demographic controls. One is the (time-invariant) ethnic fractionalization index from Alesina et al. (2003); a fractionalized society may make it more difficult, or less desirable, to implement pension systems (see, e.g., Alesina, Baquir and Easterly 1999). The other is log population size, from Maddison (2007). Although the latter is statistically significant – larger countries more often have pension systems, possibly due to the increased complexity of running distributive schemes though patronage networks in larger countries, necessitating formal policy programs such as old-age pension systems – BMR does not change much. In Model A3 we further add urbanization, which correlates with regime type and pensions, to mitigate potential omitted variable bias. We also add two variables capturing available resources for, and alternative strategies of, co-optation (and repression). These are revenues from oil, gas, coal and metals as percentage of GDP and military personnel as percentage of the population.11 BMR turns negative in A3, but remains 11

Urbanization, resource dependence, and military size are from Miller (2013). The results do not

change when, e.g., rather using natural resource income measures from Haber and Menaldo (2011).

16

insignificant. These (non-)results are robust to adding other potentially relevant controls (taken from Haber and Menaldo 2011; Miller 2013), such as civil war, British ex-colony status, literacy rates, or dependence on agriculture (Appendix Table A.2). We also controlled for trade to account for stronger incentives to adopt social policies in open economies (e.g., Mares 2005), and our results are robust. The results are retained when lagging all independent variables, substituting BMR with the continuous Polity index, controlling for many autocracies being Communist regimes, and running the models on post-1945 samples only (see Appendix).12 Models A4–A6 resemble A1–A3, but only include countries without pension system in t − 1 and have introduction of pension system in t as dependent variable. These models show that autocracies are no less likely to adopt pension systems than democracies; while BMR is always positive, the t-values range from 1.1–1.3. Thus, there is no robust evidence that democracy is systematically associated with having or adopting old-age pension systems. This corroborates H1. These results do not, however, imply that the typical “Autocratic Welfare State” is identical to the democratic. Indeed, our argument provided several indications on how they may differ. First, since autocratic coalitions often only comprise particular groups, some programs, other than pensions, may be less common in autocracies. For example, autocracies may less often implement unemployment insurance programs, because the main beneficiaries – unemployed workers – are presumably often outside the autocrat’s coalition. Instead, employment protection legislation might fulfill the autocrat’s need to insure and win over targeted groups of high income workers, without redistributing to groups posing little risk to regime security. Models B1–B5, Table 2 then also show less support for a “generalized version” of H1 applying to all types of social policies rather than pensions only. When running models resembling A3 from Table 1, but on existence of (above minimum-threshold) unemployment benefits-, maternity leave-, family allowance-, work injury-, and sickness benefits programs, democracy is positively associated with three of them: BMR is significant at 5 percent for unemployment benefits (B1), work injury (B4), and sickness leave (B5). These policies are presumably less easy (than pension programs) to use for 12

However, there are indications that democracies more often have pension systems once using the

DD measure from Cheibub, Gandhi and Vreeland (2010). This partly reflects the shorter sample (post1945) – t-values are higher (though still insignificant) for BMR post-1945. We also tested various Cox duration model specifications. Here, the regime coefficient was sensitive to minor specification changes, with many models showing no effect and some showing a positive effect of democracy.

17

Program Democracy (BMR) Ln GDP p.c. Ln population Ethnic fraction. Urbanization Size military Resource dependence Region dummies Year dummies N Countries

B1 Unemployment b/(t) 1.002** (2.45) 1.184** (2.45) 0.746*** (3.62) -1.676 (-1.25) 0.015 (0.83) 0.260 (1.38) -0.011 (-0.33) Y Y 7039 130

B2 Maternity b/(t) 0.272 (0.91) -0.209 (-0.58) 0.279 (1.26) -1.052 (-1.03) 0.024** (2.05) 0.106 (0.48) -0.004 (-0.33) Y Y 7825 136

B3 Family allowance b/(t) 0.289 (0.82) 0.272 (0.75) 0.092 (0.42) -0.787 (-0.74) 0.039*** (2.84) -0.303 (-1.42) -0.049*** (-2.58) Y Y 6724 136

B4 Work injury b/(t) 1.036*** (2.79) -0.446 (-0.91) 0.002 (0.01) -0.396 (-0.29) 0.004 (0.18) 0.347 (0.78) 0.011 (0.47) Y Y 7765 135

B5 Sickness b/(t) 0.650** (2.50) 0.002 (0.00) 0.560** (1.97) -2.390** (-2.12) -0.005 (-0.41) 0.020 (0.11) 0.009 (0.55) Y Y 7894 136

Table 2: Democracies, autocracies, and the existence of various welfare programs. Notes: *** p < 0.01, ** p < 0.05, * p < 0.10. Logit regressions with errors clustered on country, with existence of program above minimum standard as dependent variables. Maximum time series extend from late 1880s–2004.

directing resources towards autocratic winning coalitions. However, democracy is not systematically related to maternity leave or family allowance programs. These results are not predicted by our theoretical argument (but does not contradict either). However, they could relate to other features of many autocratic regimes, such as the active pro-natalist policies pursued in many autocracies to increase child births and expand the population (and, autocracies do indeed have higher population growth, see Przeworski et al. 2000, chapter 5). Our argument also indicated that autocracies, depending on how they are institutionalized, differ in incentives and capabilities to construct and commit to distributive social policy schemes. We employ data from Geddes, Wright and Frantz (2014), distinguishing regimes after who controls access to offices and policy-making decisions. One could, for example, expect dominant-party regimes to more often have pension systems than the other autocracy types (personalist, military, monarchies), due to their typically larger supporting coalitions. We thus ran models on probability of having and adopting pension systems, but substituting BMR with the Geddes et al. dummies. These models (Appendix Table A.8) have “pure” dominant-party regimes as reference category (hybrid dominantparty regimes, such as “dominant-party–military”, receive their own dummy). All autocratic regime dummies are, indeed, consistently negative. Also the democracy dummy is negative more often than not. Thus, the point estimates indicate that dominant-party regimes more often have, or adopt, pen-

18

sion programs. Surprisingly, however, personalist regimes are not systematically less likely to have pension programs than dominant-party regimes. One potential reason, alluded to above, is that some ruling parties may be so strongly institutionalized, providing clear checks on the leadership (e.g., Svolik 2012), that promises from regime leaders are credible even absent formalized pension programs. In personalist regimes, fewer institutional checks on the ruler exist, and formalized pension programs may be among the only available options for credibly committing to future distribution, even if they are costly to set up for relatively small ruling coalitions. In contrast, the results are often statistically significant at conventional levels for the military regime and monarchy dummies, also when controlling for dominant-party regimes more often being Communist.13 We also tested whether dominant-party autocracies more often have other social policy programs. While the results are not clear, dominant-party regimes, generally, seem to correlate also with such programs (Appendix Table A.11). For instance, dominant-party regimes more often have sickness benefits than monarchies and personalist regimes, but there are no significant differences on unemployment. 5.2

Do autocracies adopt less universal pension programs than democracies?

Despite similar propensities for democracies and autocracies to have pension systems, our argument indicates that these pension systems should differ in how they are structured. We expect such systems to cover more groups in democracies. Indeed, the mean Universalism-Index score for pensions is 3.4 for democracies and 1.5 for autocracies. In post-WWII samples, the respective numbers are 4.4 and 3.0. Yet, income level, or country-specific historical factors, could affect both regime type and pension features. We thus run different models, controlling both for year- and country-specific effects, to test H2. We use linear models for our baseline specifications, but in the appendix we report various robustness tests, including ordinal logit and probit specifications and logit and probit models using dummies based on various cut-offs for the index, and these provide the same result (i.e., support for H2). Models C1–C2, Table 3, report standard fixed effects specifications with errors clustered on country. C1 controls also for income level, and the point estimate indicates that going from autocracy to democracy increases the pension system Universalism Index (ranging from 0–9) with 0.5 (t=2.03), a 13

The results are retained for existence of pension programs, although not for adoption, when us-

ing the original dominant-party dummy (including hybrid regimes) from Geddes et al. as reference category. 19

Democracy (BMR) Ln GDP p.c.

C1 Fixed Effects b/(t) 0.528** (2.03) 0.333 (1.31)

Ln population Urbanization Size military Resource dep.

C2 Fixed Effects b/(t) 0.128 (0.61) -0.005 (-0.02) -0.340 (-1.22) 0.056*** (5.36) 0.025 (0.25) -0.008 (-1.17)

Lag dep. var. Year dummies Y N Countries

Y Y 10727 139

Y Y 8128 138

C3 ECM b/(t) 0.068*** (2.70) 0.051** (2.18)

-0.069*** (-9.13) Y Y 10600 139

C4 ECM b/(t) 0.052* (1.82) 0.040 (1.28) -0.009 (-0.34) 0.005*** (3.71) 0.025 (1.57) -0.000 (-0.06) -0.091*** (-9.81) Y Y 8057 138

C5 System GMM b/(t) 0.088** (2.10) 0.031 (0.78)

0.932*** (53.22) Y Y 10600 139

C6 System GMM b/(t) 0.092** (1.98) 0.017 (0.26) 0.005 (0.30) 0.001 (0.42) -0.000 (-0.01) 0.003** (2.07) 0.923*** (44.79) Y 8057 138

Table 3: Democracies, autocracies, and universalism of pension systems. Notes: *** p < 0.01, ** p < 0.05, * p < 0.10. Universalism Index for pensions is dependent variable. Errors are clustered on country. Maximum time series extend from late-1880s–2004.

modest increase. We also tested a random effects specification, and the result was very similar (see Appendix). However, BMR loses significance in C2, which adds population, urbanization, military size and resource dependence as controls. However, to further limit endogeneity concerns (for example due to targeted pension programs being conducive to autocracy), we run Error Correction Models (ECM). In addition to country-fixed effects, these models control for the lagged dependent variable (and lag the independent). Hence, C3 and C4 investigate whether regime type in t − 1 affects changes in universalism from t − 1 to t. This is, indeed, the case, both according to C3 (t=2.70) and the more extensive C4 (t=1.82).14 We also run a more complex dynamic panel model, namely the system GMM estimator from Blundell and Bond (1998), which is particularly suitable for analyzing slow-moving variables (such as regime type 14

Nevertheless, social policy programs are fairly stable. For pensions, 94% of our observations show

no change in the Universalism index. Eligibility is also extended more often than withdrawn. 90.9 percent of all changes in democracies were expansions, and 90.1 percent in autocracies. As expected from our argument, few regimes dare scale down pension benefits to groups already receiving them. These instances then tend to appear right after a new regime takes power (our data show that the relative frequency is almost three times higher in the first three years after regime inception), and likely concern groups only perceived critical to the former regime’s survival (see also Albertus and Menaldo 2012).

20

and pension programs). As for ECM, system GMM models account for country-specific effects, and include lagged dependent variables. These models, both the parsimonious (C5; t=2.10) and extensive (C6; t=1.98), basically replicate the ECM results.15 Thus going from autocracy to democracy is associated with subsequent increases in the universalism of old-age pensions. In sum, democracy is positively associated with level of universalism (fixed effects) and even more clearly with subsequent positive changes in universalism score (ECM and GMM). Autocracies thus target their pension spending towards fewer social groups than democracies do. However, on this account, our argument did not imply that the association is likely restricted to pensions; the electoral and other political dynamics in democracies may induce them to promote more universal programs across the board. To provide an overall evaluation, we construct a simple, additive Welfare State Universalism Index (WSUI), which sums the Universalism Indices for all 6 program types (thus ranging from 0–54). Although not completely robust, the result is generally corroborated also for WSUI. Appendix Table A.12 presents six models equivalent to those in Table 3 but with WSUI as dependent variable. The results are qualitatively similar to the pension results, with all models, except the extensive fixed effects model (C8), yielding significant, positive BMR coefficients. Thus, democracies have more universal social policies, in general, than autocracies. In extension, we investigated whether dominant-party autocracies have less universal pension programs than other autocracies, and we further tested models using WSUI to check for differences between dictatorships regarding overall level of universalism. The results are reported in Appendix Table A.19 (pensions) and A.20 (WSUI). We do not find solid evidence of systematic differences between the autocracy types. This may be considered a surprising null-finding not only in light of our argument, but also since the wider literature on varieties of autocracy highlight the mass support base and broader supporting coalitions of party-based autocracies (e.g., Geddes 1999; Magaloni 2006; Fjelde 2010; Levitsky and Way 2010). To be fair, the point estimates suggest that dominant-party regimes, on average, have more universal policies, including pension programs, but the coefficients 15

The results are fairly robust to lagging all independent variables one additional year. We also

tested system GMM models instrumenting for levels of the endogenous independent variable (regime type) by using lagged differences, and for differences in regime type by using lagged levels (see Roodman 2009). To mitigate the “too-many-instruments” problem, we limited the number of lags used for instrumentation to 4. The point-estimates were basically similar to the more efficient C5 and C6, but the t-values dropped to 1.5–1.6 (see Appendix).

21

are mostly non-significant. There are some exceptions; System GMM models treating regime type as endogenous report that pure dominant-party regimes have more universal pension programs than monarchies, personalist regimes and hybrid-party regimes. Nevertheless, the lack of robustness means that we cannot conclude on any systematic difference between different types of dictatorships in terms of welfare state universalism. 5.3

Do pension programs reduce risks of autocratic regime breakdown? If pension systems constitute the survival tool that we propose, they should reduce the probability

of autocratic breakdown. We thus test H3 by employing data from Geddes et al. (2014) on autocratic regime failure. These data draw on a regime definition emphasizing formal and informal rules for selecting policies and leaders, and the measure captures changes from autocracy to democracy, between different autocracy types, and even between regimes of similar type (e.g., between the Personalist, but distinct, Mobutu and Kabila regimes in Zaire/Congo). Initially, we employ a naive assumption and consider pension programs as exogenous – despite the evidence in Table 1 showing they are endogenous to several factors, and our argument indicating that fragile autocratic regimes should have stronger incentives to adopt (costly) pension systems. We run logit and probit models with regime failure as dependent variable, controlling for ethnic fractionalization, GDP per capita, population, military size and resource dependence. We also include the Geddes regime dummies, since autocracy type may affect probability of adopting pension systems, as seen above, and regime survival (Geddes 1999; Hadenius and Teorell 2007). We include log regime duration in some models; older regimes are generally at lower risk of being deposed (e.g., Svolik 2012). However, tenure length may partly stem from historical existence of pension programs, which is strongly correlated with current programs. Hence, including regime duration may induce post-treatment bias, and we thus also report models without it. These models are reported in Appendix Table A.22. Given the plausible argument on how pension programs bolster autocratic regimes through serving as a co-optation device allowing both for targeting and credible commitments, the lack of statistically significant association between pension programs and regime failure might, at first glance, be surprising. Is there no effect on autocratic survival, after all? Such a conclusion would, however, not only be premature, but incongruent with the argument H3 is drawn from. We expected autocrats to implement expensive pension programs where the regime perceives its future survival is threatened. In such situations, the regime may initi-

22

ate pension programs to shore up coalition support, or co-opt new groups. Pension programs should thus tend to exist in contexts where probability of regime survival is, initially, low. Even if pension programs should have a positive causal effect on regime survival, they may have zero correlation. Hence, we must model the “selection” of pension programs and the effect of these programs on regime survival simultaneously. There is no established duration model set-up that allows us to properly model endogenous independent variables, and we did therefore not pursue any tests using such models here. We rather employ Instrumental Variables Probit (IVProbit) models to deal with the vital endogeneity question. Since this model can be fairly sensitive to specification choices, we test different combinations of controls and instruments, and also run (more robust) 2SLS models. To produce consistent estimates of the causal effect, we must identify instruments that are strongly correlated with pension programs, and that are not directly related to regime failure. Drawing on an insight from the literature on institutions and economic outcomes (Persson and Tabellini 2003; Knutsen 2011; Huber, Ogorzalek and Gore 2012; Acemoglu et al. 2014) we construct instruments tapping exogenous regional (and global) variation in the endogenous independent variable. The underlying notion – backed up by literature on cross-border learning and other diffusion mechanisms pertaining to social policy design (e.g., Obinger, Schmitt and Starke 2012) – is that regimes residing in neighborhoods where most countries have pension programs are more likely to adopt and continue such programs, everything else equal. While some cross-country variation in adoption, as argued, stems from perceived regime threats and strategic calculation on the part of the regime, this is, of course, not the only source of variation. We aim at isolating the (exogenous) share of variation not affected by strategic adoption; Weyland (2005, 2007), for instance, highlights how various cognitive biases may make leaders more likely to adopt regionally “common models” of such policies, inducing cross-border “learning”. Further, once controlling for the other covariates (with one caveat related to neighboring regime breakdowns accounted for below), neighboring country pension programs should not directly impact on regime failure. The latter point point is important, as the exclusion restriction is conditional on the control variablesincluded, and we test models controlling for different covariate sets, for instance including time-trend/year-fixed effects, region-/country-fixed effects and regime age. Hence, our instrument – percentage share of other countries in the region (as defined by Miller 2013) having a pension program in that year – is expectedly a strong instrument and – while this can never be fully ensured – should stand a good chance of satisfying the exclusion

23

restriction. Importantly, these propositions are clearly supported by diagnostics tests on instrument strength and the exclusion restriction, which are presented below. In some models, we add share of (other) countries globally with pension programs, to pick up exogenous global trends, but this instrument is weaker.

24

25 3248 98 R 147.3 16.4

3248 98 R

148.6 16.4

0.000 (0.08)

-0.867** (-2.33) 0.724*** (5.66) -0.374* (-1.71) 0.444*** (4.39) 0.389 (1.25) -0.093 (-1.16) -0.044 (-1.25) -0.069 (-0.40) 0.008* (1.91) -0.145* (-1.86) -0.000 (-0.06) 0.095** (2.15)

D2 IVProbit b/(t)

3248 98 R+G 0.30 81.0 19.9

0.002 (0.41)

-0.735** (-2.18) 0.650*** (6.10) -0.275 (-1.39) 0.398*** (4.46) 0.479* (1.75) -0.079 (-1.08) -0.035 (-1.08) -0.090 (-0.56) 0.007* (1.80) -0.134* (-1.78) -0.001 (-0.14)

D3 IVProbit b/(t)

3248 98 R+G 0.27 80.1 19.9

0.000 (0.04)

-0.831** (-2.19) 0.731*** (5.82) -0.360* (-1.67) 0.445*** (4.42) 0.395 (1.27) -0.094 (-1.20) -0.045 (-1.30) -0.073 (-0.42) 0.008* (1.86) -0.145* (-1.86) -0.000 (-0.08) 0.094** (2.12)

D4 IVProbit b/(t)

3247 98 R+G 0.48 81.3 19.9

2.758 (1.43) 1.226** (2.00) 0.003 (0.69)

-0.720** (-2.11) 0.656*** (6.24) -0.256 (-1.30) 0.411*** (4.55) 0.468* (1.72) -0.083 (-1.12) -0.028 (-0.85) -0.096 (-0.60) 0.007* (1.75) -0.113 (-1.50) -0.001 (-0.19)

D5 IVProbit b/(t)

3247 98 R+G 0.45 80.5 19.9

-0.809** (-2.13) 0.730*** (5.94) -0.335 (-1.57) 0.454*** (4.47) 0.393 (1.28) -0.097 (-1.22) -0.037 (-1.07) -0.080 (-0.47) 0.008* (1.80) -0.124 (-1.59) -0.001 (-0.13) 0.087** (2.00) 2.644 (1.38) 1.191* (1.96) 0.001 (0.33)

D6 IVProbit b/(t)

3247 98 R+G 0.48 81.3 19.9

Y

1.750 (1.02) 0.655 (1.18) 0.017*** (3.44)

-2.481*** (-4.50) 0.182 (0.68) -0.427 (-1.64) 0.148 (0.80) 0.130 (0.47) -0.055 (-0.49) 0.152** (1.98) 0.169 (0.58) 0.010* (1.78) -0.011 (-0.13) 0.002 (0.55)

D7 IVProbit b/(t)

3247 98 R+G 0.45 80.5 19.9

Y

-2.551*** (-4.74) 0.251 (0.84) -0.501* (-1.90) 0.185 (0.91) 0.057 (0.20) -0.065 (-0.55) 0.145* (1.86) 0.181 (0.60) 0.010* (1.81) -0.020 (-0.24) 0.002 (0.55) 0.080 (1.45) 1.566 (0.93) 0.618 (1.14) 0.015*** (2.98)

D8 IVProbit b/(t)

53.3 16.4

3120 98 R

Y Y

-3.119*** (-14.37) -0.190 (-0.63) -0.475* (-1.94) 0.004 (0.02) -0.147 (-0.50) -0.000 (-0.00) 0.222*** (3.31) 0.190 (0.55) 0.013** (2.25) 0.008 (0.08) 0.001 (0.22)

D9 IVProbit b/(t)

52.6 16.4

3120 98 R

Y Y

-3.141*** (-16.21) -0.147 (-0.44) -0.520** (-2.11) 0.025 (0.11) -0.203 (-0.70) -0.005 (-0.04) 0.216*** (3.19) 0.194 (0.56) 0.013** (2.25) 0.003 (0.03) 0.001 (0.26) 0.054 (0.85)

D10 IVProbit b/(t)

3283 97 R+G 0.42 82.5 19.9

Y

0.380* (1.88) 0.127* (1.77) 0.003*** (2.98)

0.002** (2.54) -0.026*** (-2.61) 0.001 (1.32)

-0.196** (-2.50) 0.077*** (3.63) 0.075** (2.40) 0.032* (1.70) 0.132 (1.30) -0.036** (-2.34) -0.051 (-1.08)

D11 FE2SLS b/(t)

3283 97 R+G 0.37 81.7 19.9

Y

0.002*** (2.63) -0.028*** (-2.78) 0.001 (1.08) 0.028*** (4.56) 0.360* (1.78) 0.130* (1.82) 0.002* (1.88)

-0.207*** (-2.63) 0.101*** (4.64) 0.032 (0.99) 0.043** (2.24) 0.136 (1.34) -0.044*** (-2.88) -0.041 (-0.88)

D12 FE2SLS b/(t)

Notes: *** p < 0.01, ** p < 0.05, * p < 0.10. R=share other countries in region having pension program that year; G=share other countries globally having pension program that year. All diagnostic tests for IVProbit models are conducted on structurally equivalent (fixed effects) 2SLS models. The first-stage regressions are reported in Appendix Table A.23. Errors are clustered on country in IVProbit models. Maximum time series are 1946–2004.

Table 4: Second-stage IV regressions with autocratic regime failure as dependent variable and existence of old-age pension programs as endogenous independent.

N Countries Instruments Sargan p-value Cr.-Donald–Wald F-stat Weak Id. crit.value (10%)

Year dummies Region dummies Country dummies

Time trend

Region regime failures

Global regime failures

Ln regime duration

Resource dependence

Size of military

Urbanization

Ethnic fractionaliz.

Ln population

Ln GDP p.c.

Other autocracy

Personalist regime

Monarchy

0.002 (0.46)

-0.770** (-2.32) 0.643*** (5.90) -0.288 (-1.43) 0.397*** (4.43) 0.474* (1.74) -0.077 (-1.04) -0.034 (-1.03) -0.086 (-0.53) 0.007* (1.85) -0.133* (-1.78) -0.001 (-0.12)

Pension program

Military regime

D1 IVProbit b/(t)

Model

Table 4 reports the second-stage results from twelve IV specifications. D1 and D2 (adding log regime duration to D1) only use the “region-share” instrument. This instrument is, quite naturally, highly collinear with the region dummies, but also with the year dummies. D1 and D2 omit these dummies, but include a linear time trend. Both models report a clear negative causal effect (significant 5 percent) of having pension programs on regime failure. Further, the region-share t-value in the first stage is very high (> 5), and weak identification tests show the instrument is sufficiently strong.16 Yet, including only one instrument disallows overidentification tests for checking the exclusion restriction. Thus, we add the “global-share” instrument in D3 and D4. The results are retained, and the Sargan p-values (around .3) indicate the exclusion restriction holds, suggesting that we may report consistent estimates of a causal effect. However, one could still, on theoretical grounds, expect the exclusion restriction to be violated if pension programs regionally and globally correlate with incidences of regime failure regionally and globally, and regime failures abroad directly impact on risk of regime failure domestically. To exclude this causal pathway affecting our specification and results, we control for shares of (other) countries regionally and globally experiencing regime failure (in given year) in D5 and D6. This does not alter the estimated effect of pension programs on regime survival, but, as expected, the Sargan p-values increase even further, thereby increasing our confidence that the exclusion restriction holds. D7 and D8 include region-fixed effects. While this clearly weakens the first-stage t-value of our regional instrument (to, respectively, 1.1 and 2.1) due to collinearity, the absolute t-values for pensions on survival increase, to 4.5 and 4.7 respectively. We also tried substituting the time trend in D7 and D8 with year dummies, but these models did not converge. After omitting the regional and global regime failure variables, and using only the regional instrument, the IVprobit regressions ran also with both region and year dummies. These models (D9 and D10) lead to exceedingly high collinearity with the instrument in the first stage, and we thus put less trust in these results. Still, if we are to believe them, there is a very strong effect of pension programs on regime survival (t < −14). Finally, D11 and D12 are linear fixed effects 2SLS specification using both the regional and global instruments. These are conservative models, controlling for country-specific effects, and they report a causal effect of 16

The first-stage results (Appendix Table A.23) also indicate that urbanization correlates with pen-

sion programs. This coheres with our argument; urban societies present more difficult challenges to autocrats, as the concentration of people within large cities eases different opposition coordination problems. Hence, autocrats have stronger incentives to adopt pension programs in urban societies.

26

Italy 1925

Argentina 1952

Thailand 1973

China 2000

0,5 0,45 0,4 0,35 0,3 0,25 0,2

0,15 0,1 0,05 0

Figure 2: Predicted probabilities of regime breakdown (from Model D1) with 90% CIs. All covariates set at actual values, except hypotheticals on pensions (Thailand-1973 did not have pension system, but the other three did) and Italy-1925 (not coded by GWF) is coded as dominant-party regime.

pension programs significant at 1 percent. While this result does not hold when substituting the time trend with year dummies, we did find a significant effect in other FE2SLS specifications. Although the results are somewhat weakened (-1.2