Does India have subnational welfare regimes? The role of state governments in shaping social policy

ABSTRACT This article asks whether state governments in India had developed sufficient autonomy during the era of multi-party fragmentation (1989–2014) to determine and implement their own distinctive policy agendas in the field of social policy. It uses cluster analysis to look for variation in the welfare regimes of state governments within India’s federal system. Employing the metrics of (de)commodification and defamilialization commonly used in the comparative cross-national literature on welfare regimes to assess the extent to which public welfare provision reduces dependence on market forces and family structures, the paper finds that India’s states cluster in ways that suggest the existence of distinctive subnational welfare regimes rather than straightforward variation within a national welfare regime. The analysis challenges the national-level bias inherent in previous studies of comparative welfare regimes and offers insights into the operation of regional autonomy within the domain of social policy in India’s federal system.


INTRODUCTION
The central government plays a leading role in the design of national social policies within India's relatively centralized federation. At a national level, India has been categorized as having an 'informal security regime' rather than a 'welfare state regime' (Wood & Gough, 2006), and very low levels of central government expenditure on social sectors such as education and health compared with other similar economies (Dreze & Sen, 2013). However, subnational variation in social provision and human development outcomes has always been evident. State governments have historically been locations for social policy innovation, as well as being the lead administrative agencies for the implementation of many centrally designed programmes. Historically only a handful of states with more social democratic political regimes stood out for progress on reducing poverty (Kohli, 1989). In recent decades, a wider range of states have been engaged in the design and implementation of social assistance policies that are less subject to clientelistic intermediation than in the past. In the era of political regionalization that occurred alongside economic liberalization from the early 1990s, many state governments began to initiate their own social policies or layer new provisions on top of central schemes which in some cases widened coverage or deepened their generosity. This article uses cluster analysis to assess whether India's states exhibited categorical differences in their welfare regimes during the period of multi-party political fragmentation .
The analysis focuses on the period 2009-14, the second term of the Congress Party-led United Progressive Alliance (UPA) government at the national level. Between 2004 and 2014, the UPA government at the Centre introduced a rights-based policy agenda which focused on establishing legal entitlements to social provision in a bid to empower citizens to demand better service delivery. A series of social rightsto education, food, employment and informationwere established via legislation (Drèze & Khera, 2017;Harriss, 2013;Jayal, 2013;Khera, 2011;Koehler & Chopra, 2014;Ruparelia, 2013). In the early years of the UPA government, India also introduced a major anti-poverty programme, the Mahatma Gandhi National Rural Employment Guarantee Act (MGNREGA), which was intended to provide paid employment, on demand, to rural households on local public works schemes. These rights-based schemes were forms of publicly funded social assistance that were not mediated by employers or employment status. They stood apart from longer-standing forms of social insurance for the dwindling pool of workers within the formal sector.
The period of the UPA government at the Centre was one of substantial political regionalization, during which states were also able to claim credit for the successful implementation of central programmes (Tillin & Pereira, 2017). State finances were also relatively buoyant allowing them to increase their social investment if they so choose (Aiyar & Kapur, 2019). The cluster analysis conducted in this article shows that, by 2014, Indian states had started to cluster in ways that suggest the existence of distinctive subnational welfare regimes rather than straightforward variation within a national welfare regime. A distinctive cluster of states had emerged by the end of the UPA government in 2014 which combined higher levels of labour force participation with the relatively effective delivery of non-employment linked social assistance. Another small cluster of states employed more decommodifying policies (with lower overall levels of labour market participation, greater public sector employment and higher overall social sector expenditure) accompanied by a high reliance on familial structures. A third cluster represent commodifying regimes in which people are forced to rely on labour markets with high levels of insecurity and weak social safety nets.
The article offers two principal contributions. First, it offers insights into the operation of regional autonomy within the domain of social policy in India's federal system. The analysis shows that state governments had developed a meaningful degree of autonomy in the field of social policy by 2014 such that distinctive welfare regimes can be identified in a wider set of subnational cases than the existing literature has traditionally focused on. The identification of subnational welfare regimes in India in this periodacross richer and poorer states, and regardless of political alignment with the Centreindicates that states had developed different policy preferences, rather than simply exhibiting differential state capacity, levels of clientelism or political incentives to implement national programmes. Second, by identifying distinct clusters of subnational welfare regimes, the analysis challenges the national level bias found in cross-national analyses to offer a picture of the emerging modal welfare regime in India by the end of the UPA regime which combined high levels of labour commodification with publicly financed social assistance.
The remainder of the paper is structured as follows. The next section reviews the comparative welfare regime literature as a basis for developing a theoretical framework for subnational comparative analysis of welfare regimes in India. It then presents an overview of the constitutional, fiscal and political dimensions of federalism as relevant for understanding the role of states in social policy. The third section presents the cluster analysis, describes the components of the three clusters of subnational welfare regimes and the wider implications of the analysis. A concluding section considers the implications of the return of a dominant party regime at the Centre since 2014 for the role of states in the field of social policy.

WELFARE REGIMES IN LOWER AND MIDDLE-INCOME COUNTRIES: TOWARDS A MULTI-LEVEL PERSPECTIVE
The emergence of Western European 'welfare states' was almost synonymous with a particular form of 'welfare capitalism' that had its heyday in Western Europe in the decades following the Second World War. Their purpose was defined in terms of the 'decommodification of the status of individuals vis-à-vis the market' (Esping-Andersen, 1989, p. 21), achieved when 'service is rendered as a matter of right, and when a person can maintain a livelihood without reliance on the market' (p. 22). Multiple patterns of welfare capitalismor constellations of relations between states, markets and familieswere identified. It is now commonly acknowledged that distinct clusters of welfare regimes are not only a phenomenon of late capitalist development in advanced industrial countries. Cross-national comparative research has revealed systematic patterns of divergence in the relationships among state, markets and labour across lower and middle-income countries too. These indicate the existence of distinctive approaches to social policy even where labour markets are characterized by high degrees of informality and in which social policy provisions do not approximate more fully fledged 'welfare states' (Abu Sharkh & Gough, 2010;Gough & Wood, 2004;Haggard & Kaufman, 2008;Holliday, 2000;Rudra, 2007).
Explaining divergence in the welfare regimes of less developed countries in the later 20th century, Rudra (2007), like other literature on social policy regimes in East Asia (Holliday, 2000;Kühner & Nakray, 2017), describes social policy strategies that promote human capital formation via education/health investment as 'productivist' as opposed to 'protective' social policies which protect workers through a mixture of worker protections and reliance on public-sector employment. Rudra argues that 'productivist' approaches are found in countries with greater historic exposure to international market competition via export-led industrialization, in contrast to import-substituting industrializerssuch as Indiawith more closed economies and higher worker protections.
Welfare regimes in developing countries are, however, not only shaped by their relative dependence on markets versus public policy. Given the high level of informality in most developing countries, non-market and non-state actors such as family, community and non-governmental organizations play a pronounced role in supporting livelihoods and coping with risk. For this reason, Gough and Wood (2004) introduced the concept of 'informal security regimes' as a form of welfare regime in which the domains of market, state, family and community are porous, compromising the ability of the state to act independently to compensate for risks generated by dependence on markets or to challenge patterns of hierarchy or dependency built into social relations (Wood, 2004). Markets are likely to be shaped by patterns of domination and exclusion based on ascriptive identities, such as caste and religion in the Indian case (on social structures of accumulation in India, see Harriss-White, 2003). These informal security regimes, in which Gough and Wood group India, contrast with 'welfare state regimes' in which 'people can reasonably expect to meet (to a varying extent) their security needs via participation in labour markets, financial markets and the finance and provisioning role of a "welfare state"' (Gough, 2004, p. 33).
The role of families as an element of the welfare mix is closely related to the sexual division of labour and the place of women as workers and/or caregivers. Feminist scholars have long argued that traditional welfare state analyses paid too little attention to gender as an organizing principle, and that welfare regimes should be judged not only through the lens of decommodification, but by the extent they engage in defamilialization (Lewis, 1992;Orloff, 1993;Sainsbury, 1999).
Familialist regimes are defined as those which place an 'emphasis upon the family as the primary locus of welfare provision by way of intra-and inter-generational mutual aid' (Estévez-Abe et al., 2016, p. 302). Defamilialization entails enabling female participation in the labour market and recognizing the value of unpaid caring work. Most research on family policy or defamilialization as a component of welfare regimes has focused on advanced industrial contexts, with Southern European and East Asian welfare regimes typically identified as the most heavily familialist (Estévez-Abe et al., 2016;Saraceno, 2016). Franzoni (2008) offers one of the few attempts to produce an integrated typology that addresses interactions between commodification, decommodification and defamilialization in welfare regimes characterized by labour market informality. Looking at patterns of social welfare across Latin American countries, Franzoni identifies three clusters: state productivist regimes (where social protection is used to encourage labour market participation), state protectionist regimes (with high levels of labour commodification, accompanied by protections for formal sector workers) and familialist regimes (marginal state provision, with welfare provided by family/community) (Franzoni, 2008;Rodriguez-Loureiro et al., 2020).
Most of the analyses summarized so far focus on employment-linked social insurance when they analyse worker protections. They have not typically included the substantial programmes of state-financed social assistance to reach labour market 'outsiders' which were introduced by many emerging economies as their growth took off in the 2000s (Barrientos, 2013;Garay, 2016;Tillin & Duckett, 2017). As Seekings (2017, p. 367) argues, many developing countries have demonstrated that it has been 'surprisingly easy' for states to expand their capacity to construct more inclusive, pro-poor welfare states through the expansion of reasonably effective, non-contributory social assistance programmes such as cash transfer or social grant programmes. Social assistance became more important as traditional forms of trade unionism and corporatist state-market-labour relations lost their pre-eminence as mechanisms to achieve the social regulation of markets (Agarwala, 2013). In India, the expansion of social assistance was linked to the adoption of the rights-based approach to social policy in which non-employment linked social assistance programmes were recognized via legislation as permanent, statutory rights or entitlements of citizenship, rather than patronage bestowed by a selective state.
This review of the cross-national literature suggests the principal components for a comparative analysis of welfare regimes in the context of labour market informality are the relative degrees of (de)commodification and defamilialization, and that in capturing the nature of decommodification it is important to pay attention both to patterns of employment (some of which are linked to worker protections and social insurance), and to non-employment linked forms of social assistance.
The literature on comparative welfare regimes has advanced scholarly understanding of categorical differences in welfare regimes among lower and middle-income countries, but its insights have not been widely examined at the subnational level. Most existing studies of comparative welfare regimes compare national cases regardless of the substantial territorial inequality within countries. This national level bias has produced misleading categorizations of welfare regimes (Giraudy & Pribble, 2019;Singh, 2019). 1 It is only very recently that scholars have begun to ask whether, by shifting our level of analysis, we might identify the existence of distinctive subnational welfare regimes that indicate meaningful variation in policy approaches to the social regulation of markets at a level below the national state (see Ratigan, 2017, on China). An improved understanding of the existence and nature of variation in subnational policy regimes holds out the prospect of improving our categorization of social policy regimes at the national level too. In this spirit, Giraudy and Pribble (2019) propose a territorial Gini index of inequality in order to adjust measures of welfare state universalism to account for subnational inequality in Latin America. This multilevel perspective offers important refinements to the conceptualization of national welfare regimes.

FEDERALISM AND SOCIAL POLICY IN INDIA
In order to ask whether states in India could constitute distinct welfare regimes, we need first to understand how social policy is designed, financed and implemented in India's federal system. 2 This section will look at the extent to which states have sufficient autonomy to establish and maintain distinct patterns of social regulation of markets.
Constitutionally, India's states have substantial responsibilities in the realm of social policy, although these are shared with the central government. In India's 1950 constitution, health and education were both listed as state subjects, with social security (including labour welfare and employment-based social insurance such as pensions, provident funds, health insurance and maternity benefits) on the concurrent list over which the Centre and states share jurisdiction (Tillin, 2021). Education became a concurrent subject in 1976. Since the 1980s, the central government has become more active in the field of social policy, implementing many 'centrally sponsored schemes' even in areas that fall on the state list. These centrally designed schemes which include many of the new rights-based policies introduced under the UPA government often require matched funding by the states. States play a central role in implementing most of these centrally designed programmes, and they are also able to initiate their own social policies or layer new provisions on top of central schemes which can widen their coverage or deepen their generosity. While India's central government plays a leading role in the design of national social policies, there is substantial territorial unevenness in delivery and outcomes at the state level. Thus India, under the UPA, stood in contrast to some other federal countries like Brazil where the role of governorsat the meso-levelin implementing social policies had been reduced (Fenwick, 2016).
In fiscal terms, India's state governments are responsible for over 70% of total social sector expenditure. As Figure 1 shows, the states' share of total expenditure steadily increased during the second term of the UPA government  reaching 72% by 2013-14. The increase in state-level social expenditure is likely to have been driven by a number of factors including growth in many states' tax revenues as economic growth took off in the mid-2000s, as well as the increasing electoral salience of welfare policies in a period of political regionalization. As Table 1 shows, there is wide variation among states in the composition of their budgets, but many statesincluding some traditionally poorer stateswere in the fiscal position to make decisions about whether to increase their social expenditure. 3 Beyond constitutional provisions and the shape of fiscal federalism as it pertains to the social sector, the contours of central and regional authority are heavily influenced by the composition of the party system. As Saxena, Sharma and Swenden argue in the introduction to this special issue (Saxena et al., 2021, in this issue), federal relations are shaped by partisan dynamics and in particular the constellation of political power at the Centre. Thus, as with other areas of policy, the relative authority of the central and state governments over social policy has altered over time. This paper focuses on a period of multi-party fragmentation in which the party system was more 'state-friendly' than in the periods that preceded or have followed it. The conclusion will briefly reflect on the post-2014 situation since the return to single party dominance at the Centre. Politically, social welfare became an important theme in elections in many states from the mid-1990s onwards. As Figure 2 shows, the 1990s and early 2000s saw a sustained increase in the position of regional parties ('other seats') in the Indian Parliament. Between 1996 and 2014, regional parties held a larger number of seats than either the Congress Party or BJP. 4 Political regionalization and economic liberalization worked hand in hand to put state governments on the frontline of responding to the contradictions of widening inequality amidst rapid economic growth. In this period, state-level politics acquired an autonomous logic, distinct from that of national politics (Yadav & Palshikar, 2008, 2009. The social bases of political parties shifted in many parts of India because of political assertion by lower castes and classes, and the consolidation of new regional parties. Coalition government involving national and regional parties became a norm at the Centre. In this context, by the 2000s, chief ministers of state governments had become increasingly able to claim electoral credit in the eyes of voters even for centrally designed and financed programmes such as the MGNREGA (Tillin & Pereira, 2017).

METHODOLOGY, DATA AND INDICATORS
Having established the theoretical expectation that state governments had sufficient autonomy in this period to establish their own social policy priorities, and presented a framework for analysing welfare regimes, this section sets out the approach to cluster analysis and presents the results. Cluster analysis allows for an examination of statistical patterns in data. It helps to organize multivariate data into subgroups in order to detect internal patterns within the data (Everitt et al., 2011). Rather than looking for relationships among variables, the method looks for similarities between cases (Rudra, 2007, p. 385). In line with the conceptual discussion of welfare regimes in lower and middle-income countries, and building on earlier cross-national comparative exercises, the cluster analysis that follows uses publicly available data to create indicators for three principal dimensions: commodification, social policy and defamilialization. It also includes two indicators of welfare regime performance. An overview of the indicators, data sources and years is given in Appendix A in the supplemental data online.
The six commodification indicators focus on the level and structure of workforce participation and the remuneration of workers. Total labour force participation and female labour force participation provide a measure for the overall reliance on labour market participation. Data on the proportion of self-employed and salaried workers capture the degree of formalization of labour markets. The final two indicators (gross national product (GNP) per capita, and poverty) provide proxies for wage levels.
The six social policy indicators focus on the extent and coverage of publicly financed, nonemployment-linked social assistance as well as overall social expenditure. The two largest components of social assistance expenditure in India, by some distance, are the MGNREGA and a subsidized food programme (Public Distribution System -PDS) which provides subsidized wheat, rice and some other commodities for purchase via fair-price shops. These accounted for approximately 33% and 49% of total social expenditure, respectively, in the late 2000s (before the further expansion of the PDS under the National Food Security Act (NFSA), 2013) (Dutta et al., 2010). The first two indicators offer a measure of the coverage of food subsidies via the PDS both among the general public, and the poorest quintile. The total percentage of the population in a given state who purchase rice or wheat from the PDS is based on calculations by Himanshu and Sen (2013) using National Sample Survey (NSS) data . This measure provides an insight into whether a state has taken the policy decision to expand coverage beyond the narrow 'below poverty line' population funded by the central government (until the expansion under the NFSA which came into force after the period under analysis). The third indicator is a measure of the number of person-days generated through the Mahatma Gandhi National Rural Employment Guarantee Scheme (MGNREGS) per poor person.
The remaining social policy indicators provide measures of total social sector expenditure per capita, 5 as well as total public expenditure on wages and salaries (both directly correlated with decommodification) and total private school enrolment (inversely correlated). Unfortunately, data on private out-of-pocket health expenditure from national health accounts has missing data for some major states and so these cannot be included in the cluster analysis.
The defamilialization indicators are fewer, again reflecting data limitations. It was not possible, for instance, to find reliable data on patterns of unpaid work. The indicators include the proportion of women in reproductive years who are workers. Levels of public childcare provision are very low across India. Variation in this indicator is therefore likely to reflect whether households are more dependent on female unpaid labour for caring for children or elderly members. It may also relate to overall levels of economic vulnerability. Two indicators on the population under 14 and over 60 years provide controls for the level of demand for caregiving (this follows Franzoni, 2008, p. 79).
Finally, to provide measures of overall welfare outcomes (beyond the performance of social assistance programmes), two performance indicators on infant mortality and literacy are used. In contexts where social policy expenditure can be subject to corrupt or clientelistic intermediation, the inclusion of a number of outcome variables provides further evidence of policy prioritization (see also Pribble, 2011).
For each indicator, data for the final year of the UPA government (2013-14) are used where available, otherwise the nearest available year for the particular data set. Data across the 19 largest Indian states for which complete data sets are available for each indicator are analysed.

CLUSTER ANALYSIS RESULTS
The number of clusters is not known in advance, so a hierarchical clustering method is most appropriate because it allows the number of groupings to be driven by the data itself. The hierarchical agglomerative cluster technique in SPSS, using the Euclidean distance to calculate the distance between units, is used. Euclidean distance measures the literal distance between two objects (Romesburg, 2004, p. 12). At the beginning of cluster analysis, each case is treated as being in a separate cluster. Each step in agglomerative cluster analysis will merge the two most similar clusters, reducing the number of clusters by one each time (Romesburg 2004, p. 15). The analysis requires the specification of criteria to be used in establishing the similarity between clusters. This study employs the approach of other similar studies (Pribble, 2011;Rudra, 2007) and uses the minimum variance or 'Ward's' method, which optimizes the minimum variance within clusters and joins together groups that produce the lowest increase in the error sum of squares. This linkage technique is considered to produce the truest clusters among alternative methods (Blashfield, 1976). These clustering techniques produce a hierarchical tree diagram or 'dendrogram' which graphically represents the relationship between cases.
The dendrogram, as shown in Figure 3 and summarized in Table 2, reveals a clear pattern among states. The first-level grouping separates states into two wide clusters of richer and poorer states. The cluster of richer and poorer states are then each divided into two clusters in ways that are suggestive of more categorical differences in their welfare regimes. Appendix B in the supplemental data online provides the data for each indicator organized by cluster. States within cluster 1 all share certain characteristics. First, they have above-average per capita income and relatively low levels of poverty: these are collectively the richer states in India. Second, all states in the cluster have above-average levels of salaried workers and below-average levels of self-employment suggesting that they have a higher level of labour market formalization than cluster 2. Third, they have a similar demographic profile with below-average population under 14 and above-average in the over-60 population, suggesting that they are further advanced in the demographic transition than cluster 2. However, despite these commonalities, there are two clear clusters among the richer states.
Cluster 1a forms what is described here as a decommodifying familial cluster. The states in this cluster (Kerala, Maharashtra, Uttarakhand and Haryana) have the lowest levels of total labour force participation of all states in Indiaan average of 515 of 1000 people. As a cluster, they also have the smallest proportion of women in reproductive years in the workforce. Compared with other states in cluster 1, they have a considerably higher level of public sector employment (as reflected in the data on public expenditure on wages and salaries). They also have the highest overall levels of social sector expenditure per capita. The inclusion of Keralathe poster child of subnational social democracyalongside Maharashtra and Haryana in this cluster may appear surprising, and it must be stressed that the performance of the states in this cluster on the social assistance variables displays internal heterogeneity. Kerala and Uttarakhand have high levels of PDS coverage and above-average MGNREGA person-days per poor person, while Haryana and Maharashtra have low levels of PDS coverage and MGNREGA work. Thus, overall, the states in cluster 1a have a relatively more decommodified labour market and stronger reliance on public employment and regulation of employment (marking them out from other richer states). This somewhat more decommodified structure of labour market participation is supplemented in Kerala and to some extent Uttarakhand, but less so in Maharashtra and Haryana by a strong commitment to social protection. Overall, the cluster has the highest level of human development outcomes as measured by infant mortality and literacy rates.
The second grouping, cluster 1b, forms a group of states that are more commodifying with respect to labour market policies, but supplement a more pro-market approach with publicly financed social assistance. This cluster can be described as commodifying with protection. These states (Andhra Pradesh, Punjab, Karnataka, Tamil Nadu, Gujarat and Himachal Pradesh) have the highest overall level of labour force participation of all states (586 per 1000 people), as well as the highest levels of female workforce participation. But this high level of labour force participation is accompanied by investment in social assistance. These states collectively have the highest levels of PDS coverage among the general population and the poorest quintile, and the highest number of MGNREGA person-days per poor person. While there is some internal heterogeneity among this cluster too, most states are consistently above the national average for social assistance variables. Only Punjab and Gujarat stand out for having very low coverage of PDS and MGNREGA, and may thus be classified more as straight commodifying states.
States within cluster 2 are collectively poorer than the first cluster and share certain other characteristics. They have higher levels of labour market informality (reflected in low levels of salaried employment and above-average rates of self-employment) and lower levels of public sector employment than cluster 1. They have very similar levels of overall labour market participation (which are not far from the levels seen in cluster 1b). They also share a demographic profile with above-average population below age 14 and below average over-60 population. But again, within this cluster of poorer states, there are two clear subgroupings. Cluster 2a contains a group of states (Chhattisgarh, Rajasthan, West Bengal and Odisha) that combine relatively high levels of commodification with investment in social protection. They share important characteristics with cluster 1b, and they are therefore also labelled commodifying with protection. Their cluster average for PDS coverage and MGNREGA workdays is approximately the same as the national average, as is their per capita social expenditure. They are the most defamilializing in terms of the proportion of women of reproductive age in the workforce (with Chhattisgarh and Rajasthan standing out as having almost the highest rates of any states in the country 6 ). Overall, this is a subcluster of states that use social assistance more effectively alongside commodification. Their welfare mix is more similar to cluster 1b than to the other poor states. 7 The final cluster of states, cluster 2b, is labelled here as commodifying, insecure. These states have the highest levels of self-employment and lowest salaried workforce of all states, they are considerably below national average in terms of their social sector expenditure and in terms of their PDS and MGNREGA performance. They have the highest levels of poverty and lowest levels of per capita income. This is a picture of high levels of precarity in which people are forced to commodify their labour, but where labour market participation is doing less to raise income and reduce poverty than in cluster 2a, and social assistance is not functioning effectively as a safety net.

DISCUSSION
This analysis demonstrates that India's state governments do cluster in distinctive ways that are suggestive of different types of welfare regime. Furthermore, it suggests that states cluster in ways that offer important refinements to the traditional conceptualization of India's national welfare regime. Depending on how they define welfare regimes, previous studies have categorized India either as a 'protective' welfare regime which selectively protects individuals from reliance on the market (Rudra, 2007) or as an 'informal security' welfare regime with 'poor levels of welfare coupled with low public commitments' (Wood & Gough, 2006). By scaling our analysis down to the subnational level, we have identified that there are now a substantial number of states (both rich and poor) which combine more commodifying or pro-market labour policies with publicly financed social assistance. This commodifying with protection cluster accounted for the largest number of states by the end of the UPA government.
The comparative subnational analysis also offers a refinement to previous conceptualization of subnational welfare regime types in India. The type of welfare regime that has been developing in these commodifying with protection states looks very different to the social democratic 'Kerala model' (and historically West Bengal model) that has loomed so large in subnational analyses of welfare or poverty reduction in India to date (Kohli, 1989;Sandbrook et al., 2007;Singh, 2011Singh, , 2016. This social democratic model has hinged on the presence of a left-of-centre political regime with a decentralized administrative structure, as well as a more cohesive subnational identity or what Singh (2011) calls a sense of 'we-ness'. By contrast, the large clustering of states that have commodifying with protection welfare regimes represent a wider variety of regional party systems and may be the more likely path to further improvements in social welfare than the social democratic route. The policy model adopted in these states also suggests that the dichotomy sometimes posed in India between an employment-led versus 'entitlements'-led approach to poverty reduction is a false one (the arguments, respectively, of Bhagwati & Panagariya, 2014;and Dreze & Sen, 2013 have often been presented as representing such a dichotomy). Many states combine both strategies. The commodifying with protection model of welfare provision does, of course, reflect the shift in national social policy since economic liberalization and especially under the UPA government since 2004. But in many cases states had their own social assistance programmes that predated the national policies introduced by the UPA or expanded upon central policies. Furthermore, subnational policy innovation informed the development of the new central government social policies under the UPA.
The ability of state governments to prioritize social assistance alongside economic growth in the 2000s did not simply reflect underlying state capacity. States with very similar levels of administrative capacity made different choices. This is exemplified in the different paths adopted by Chhattisgarh and Madhya Pradesh which were only bifurcated into separate states in 2000 and have since adopted markedly different approaches to social assistance (Chhotray et al., 2020;Heath & Tillin, 2018;Tillin et al., 2015). The choices made by state governments about how to administer or prioritize social assistance programmes were also not driven by patterns of partisan congruence between the state government and central government (contrary to patterns observed in Latin America in a similar period ;Niedzwiecki, 2018). Many of the states in clusters 1b and 2a were ruled by political parties opposed to the UPA government at the Centre.
However, as the final cluster of commodifying insecure states in the final cluster 2b suggest, we should not disregard the question of state capacity entirely. These are some of India's most populous states, many of which share a key defining feature: an enduring history of upper caste dominance which came under more sustained challenge during the 1980s and 1990s by lower caste assertion (especially in the case of Bihar and Uttar Pradesh), or tribal assertion (in the case of Jharkhand). Previous research has shown that state-level differences in poverty reduction in India are influenced by the extent to which lower castes have been represented within regional political regimes (Harriss, 2002). In the large states of Bihar and Uttar Pradesh, new regional parties representing lower castes and Dalits only came to power from 1990 onwards. In both states, these parties set out to radically alter the social composition of the bureaucracy, as well as political class. This social revolution impacted state administrative capacity (Mathew & Moore, 2011;Witsoe, 2013). Similarly, in Jharkhand, decades of tribal struggles for a separate state from Bihar to counter 'internal colonialism' of this natural resource rich region, created extreme political fragmentation which undermined state administrative capacity after the state was created in 2000 (Tillin, 2013). A history of upper caste social and political dominance is true also for the states in cluster 2a, however these states have not seen such strong political mobilization by either tribal or lower caste communities. In states such as Chhattisgarh and Odisha, state-level political leaders (often with upper caste backgrounds themselves) have used social assistance as a means of bolstering their political legitimacy with the rural poor. The longer term implications for social outcomes of these divergent political paths of states with legacies of upper caste dominance remain to be seen.

CONCLUSIONS
This article has demonstrated that during the period in which India operated as a more decentralized federation , state governments had the autonomy to shape their own subnational welfare regimes. They did so in ways that produced distinctive configurations of relationships between the state, markets and family/community actors in the post-liberalization era. The article has shown that a distinctive cluster of statesboth rich and poorwere evident by the end of the UPA government in 2014 which combined commodification (labour market reliance) with the relatively effective delivery of social assistance. State governments in this period, whether opposition ruled or ruled by Congress/allies, had incentives to invest in the delivery and expansion of social assistance, even when funded by the central government, because they were able to claim credit for such policies in the eyes of voters.
As we conclude, it is important to underline that federal relations may change in ways that impact the policy autonomy of state governments. Since 2014 and the re-emergence of a oneparty dominant system under the BJP, the incentives faced by opposition-ruled state governments are changing. Voters are now more likely to attribute credit for new social policy initiatives and some older social programmes such as the MGNREGA to the central government as opposed to state governments (Deshpande et al., 2019). Furthermore, the central government has increasingly turned to delivery mechanisms that enable it to bypass state governments whether by direct communication to the district-level, through the use of technology to make direct cash payments, or through reliance on central agencies.
At the time of writing, there are insufficient public data to enable a full comparison of the situation prevailing today with that in 2014 across the indicators used in this analysis. There are reasons to expect that partisan dynamics may exert greater influence on the subnational political commitment to implement centrally designed social policies since 2014, or that central government initiatives will simply overshadow state government initiatives if subnational fiscal and policy space contracts substantially. However, our definition of 'welfare regimes' as constellations linking markets, family and the state involve phenomena that are more and less susceptible to change by state (and central) governments over the shorter term. There are longer temporal and societal dimensions that shape patterns of labour market participation and the status of women. Improvements in health and education outcomes, as well as labour market structures, are typically slow-moving whereas social assistance policies can be implemented and produce results in shorter timeframes. Subnational welfare regimes establish their own path dependencies in which shorter term policy preferences interact with longer term patterns of economic development to produce the distinctive clusters of states identified in this article. Thus, in an era of recentralization, some variation in subnational welfare regimes is likely to persist even if the fiscal and political autonomy of states is squeezed in such a way as to reduce their policy initiative.
2. An investigation into the existence of subnational welfare regimes in India, and other developing countries, rests on the assumption that it is meaningful to look for policy variation of a programmatic kind, rather than seeing clientelism or targeted transfers to political clients as over-determinant of policy outcomes. There is by now sufficient evidence of approaches to social policy design and implementation that prevent clientelistic intermediation, to warrant such an assumption even if clientelistic forms of politics have persisted alongside more programmatic policy approaches (e.g., Elliott, 2011;Maiorano, 2014;Manor, 2010;Wyatt, 2013).
3. An increase in central tax revenue shared with the states in 2015 (from 32% to 42%) also theoretically enhanced the proportion of untied funds received by state governments. 4. Since 2014, the BJP has emerged as the dominant party, signalling a return to a one-party-dominant system. Although while dominant at the national level, the BJP has suffered a series of defeats at the state level, indicating that political regionalization remains an important force (Aiyar & Tillin, 2020;Jaffrelot & Verniers, 2020;Schakel et al., 2019;Schakel & Swenden, 2018). 5. Some studies opt for social sector expenditure as a proportion of total expenditure in order to avoid variation based on wealth rather than policy priorities (e.g., Ratigan, 2017;Rudra, 2007). To test for this, the cluster analysis was again run using ratio rather than per capita data on social expenditure. There was no change in the clustering. 6. In these poor states, higher female labour force participation is likely to be as much a marker of the feminization of the workforce as a question of defamilialization driven by public policy decisions. The author is grateful to an anonymous reviewer for this observation. 7. West Bengal may approximate more of a historically decommodifying state that has since adopted social assistance. Its overall rates of labour force participation are among the lowest in the country. This may reflect the longer term legacy of left-front rule in West Bengal. ORCID Louise Tillin http://orcid.org/0000-0002-2788-9909