Which Wheel Gets the Grease? Constituent Agency and Sub-national World Bank Aid Allocation

Abstract

Questions of aid allocation have long focused on discerning the motivation of development donors. Less attention has been paid to the interests and agency of recipient state governments and even less to the interests and agency of constituencies within those states. An implicit assumption is often that the ‘poor’ either passively receive the patronage of their benefactors or they don’t. In this paper, we instead suggest that depending on the motivation of a donor, their sensitivity to needy subnational constituencies in aid allocation also depends on the political empowerment of those groups. In particular, we take advantage of the unique socio-cultural structure in India to examine if the political agency of scheduled castes and tribes (SC/STs) can explain patterns of district-level allocation of World Bank education aid. Using district-level data on a multi-year World Bank education program, district-level proportions of SC/ST population and of members of parliament we identify poor, but empowered, constituencies. We find that SC/ST districts receive more aid, even when controlling for baseline poverty and educational performance, but that these results are strongest when these districts are politically empowered. Our findings suggest that while donors may indeed respond to recipient needs, those recipients who also speak loudly for themselves fare better, highlighting the importance of constituent agency.

1. Introduction

How foreign aid is allocated has been the subject of a vast literature, spanning decades of research. The historical distinctions have been between ‘recipients’ needs’, ‘recipients’ merit’, and ‘donors’ interests’ with a focus on understanding differences in cross-national allocation patterns (McKinlay & Little, 1978). Recent work has not only seen more nuanced theoretical development (Bermeo, 2017), but also a move to understanding sub-national allocation patterns (Briggs, 2014, 2017, 2018; Jablonski, 2014; Nunnenkamp, Öhler, & Andrés, 2017). However, this work has consistently found no evidence of pro-poor allocation (Briggs, 2017, 2018; Öhler et al., 2019). Meanwhile, this literature has almost exclusively focused on the supply side of foreign aid allocation – the motivations, tactics, or methods of donors and/or their governmental or non-governmental agents. Less attention has been paid to the demand side of foreign aid allocation – if how and why the ultimate aid beneficiaries are able to effectively influence their own aid allocation.

This paper proposes that if donors are sufficiently pro-poor in their motivations, they will allocate aid to needy constituencies, especially when those constituencies are politically empowered. Politically empowered development constituencies can serve to crystalise demand and reduce search costs but can also deliver both input and output legitimacy to donors. Additionally, they can credibly challenge donors with an ex post reputational cost of neglecting the constituency. Vocal constituencies who are not served by aid distributions can make the ‘failings’ of donor institutions known to their stakeholders and broader publics.

To examine these dynamics, we focus on the allocation of the World Bank’s District Primary Education Project (DPEP) in India from 1994 to 2001. Due to organisational structures of multiple and collective principals, scholars have long argued that international organisations (IOs) may have sufficient ‘independence’ or ‘agency slack’ that can render space for pro-poor behaviour (Nielson & Tierney, 2003). This makes the World Bank a useful donor for examining our theoretical propositions. Likewise, focusing on India allows us to avail of the country’s salient socio-cultural class structure and officially designated scheduled castes and schedules tribes (SC/STs). These societal distinctions enable us to introduce a crucial distinction into our analyses by letting us separate political constituencies within the more general ‘poor’. Moreover, as the country with largest absolute numbers of impoverished citizens in the world, India is also the largest recipient of aid from the World Bank and the distribution of that aid has varied widely across India’s expansive geography, allowing for reasonable identification of subnational allocation patterns.

Our focus on the education sector also yields benefits for this analysis. As aid to education is unlikely to have any immediate geo-strategic or economic payoff, it is a most likely sector for allocation that is not self-interested for the donor. Likewise, unlike economic infrastructure or, indeed, even commercial ventures, there are few, if any, a priori constraints on the geographic allocation of education aid. Most pragmatically, we can identify allocations of the DPEP project at the district level, providing sufficient variation for our analysis. These features allow for a focused analysis of the relationship we propose.

In the following sections, the paper first develops a theoretical framework of sub-national aid allocation that depends on the motivation of the donor actor, the interests of recipient governments, and the political empowerment of the (potential) recipient constituencies. The paper then briefly outlines India’s education sector development and the World Bank’s role therein before turning to an analysis of the World Bank education programs across 593 districts which considers measures of district-level poverty, the proportion of SC/ST population, and the SC/ST status of members of parliament. The empirical study finds that districts with increased SC populations do receive more education aid from the World Bank beyond that expected to generally ‘poor’ districts, but that the allocation is magnified substantially when a district also has a SC/ST member of parliament in government. Notably, this finding suggests that it is not necessarily the ‘neediest’ that received aid, but instead the needy who were also able press their interests, that is the ‘squeaky wheels.’

2. Constituency agency and subnational aid allocation: theoretical foundations

When discussing determinants of aid allocation from donors, scholars normally concentrate on two concepts, ‘motivation’ and ‘strategy’. The motivation debate seeks to disentangle the balance of egoism and altruism in aid allocation (Acht, Mahmoud, & Thiele, 2015; Berthélemy, 2006; Hoeffler & Outram, 2011; Younas, 2008). However, this literature focuses almost exclusively on supply-side motivations, recipient countries’ needs and/or strategic importance to potential donors. A few exceptions consider the demands and/or interactions with recipient state governments – who are largely presented as unitary actors (Swedlund, 2017a, 2017b).

Only recently has scholarship turned to examining subnational aid allocation patterns. These studies have concentrated on whether aid goes to poorer and more vulnerable places (Barrett, 2015; Briggs, 2017, 2018), regions with unique political preferences (Albertus, 2015; Briggs, 2014), political violence (Bezerra & Brainthwaite, 2016), and/or needs (Öhler & Nunnenkamp, 2014). Notably Briggs (2017, 2018) finds no evidence that the World Bank and the African Development Bank target their aid to the poorest regions of recipient countries. Briggs’ (2017, 2018) theoretical discussion rests largely on issues of aid bargaining and donor control, suggesting that if aid doesn’t flow to the poorest regions, this is evidence that donors have lost control to recipient’s domestic political economy considerations. Nunnemkamp et al. (2017) also explore the allocation of the World Bank’s projects in India, disaggregating their analysis by sector. Besides considering sub-national need, they also entertain the ‘merit’ of sub-national administration. They find little overall evidence of needs-based allocation, although they do find evidence of sector-specific targeting, specifically in the health, water/sanitation, and transportation sectors. However, these studies, like the country-level literature above, still largely focus on supply-side dynamics. To the extent recipient political economy is considered, it is as an either/or – donors either have control or they lose it to the black box of local political considerations. In contrast, Swedlund (2017a) presents a more nuanced idea that the aid allocation and delivery processes represent a carefully choreographed ‘development dance’. Rather than an all-or nothing logic, this work is suggestive that subnational allocation outcomes are ultimately the result of a compromise donor and recipient interests. Yet, even here, the focus is primarily on donor interactions with the recipient country governments.

We expand upon the insights above by suggesting that the political agency of the targeted constituencies, themselves, can help explain subnational aid allocation. Rather than conceptualising the ‘development dance’ as a two-partner engagement, sub-national aid allocation is likely to depend on three-way dynamics between the donor, the recipient government, and the targeted constituencies whose interests may or may not be adequately represented in government. Indeed, subnational constituencies most in need are likely those that are also marginalised by their society and/or government. Thus, relying on constituent governments to advance the needs of these constituencies is a dicey proposition. Beyond this, there is likely to be heterogeneity even amongst the needy constituencies. In many countries, ‘the needy’ will constitute numerous, diverse, groups and may, indeed, vie amongst themselves for resources. We propose that when these constituencies are sufficiently politically empowered,¹ they may advance their own interests vis-à-vis a pro-poor donor in two ways.

We first suggest that these constituencies can help their own cause by reducing the search costs of the donor. Similar to the ‘domestic expert’ of Fang and Stone (2012) formal treatment of IO decision-making, domestic constituencies may have private, or at least asymmetric, information about their own needs. Need-based metrics, such as income or wealth levels, health or education indicators, or equality measures may be difficult and expensive for donors to obtain, especially at a sufficiently localised and sub-national level that facilitates targeting community-based interventions (Galasso & Ravallion, 2005). Empowered constituencies may have private information and will be strongly incentivised to present their ‘needs’ case, and this can reduce the costs for the donor in identifying whom to target.² Beyond this, even if public metrics exist, they may be crude and/or insufficient to determine true need. Marginalised constituencies may provide donors with qualitative or narrative evidence of need.

Indeed, recognition of this imperfect information has largely driven the World Bank’s ‘Community-Driven Development’ (CDD) initiative (Wong, 2012). CDD Programs have attempted to directly target local need by leveraging local identification of priorities.³ However, numerous studies of CDD initiatives have found that they tend to be captured by local elites (Mansuri & Rao, 2004; Platteau, 2004) and that they may not have ultimately effectively targeted the poor (Saguin, 2018).

Second, empowered constituencies can serve as a gate keeper for donor access and input and output legitimacies. Altruistic donors are at pains to demonstrate to their stakeholders that their efforts do indeed have a ‘pro-poor’ focus (Younas, 2008). This can often come through endorsement of a needy constituency itself.⁴ Politically empowered, vocal, and/or visible constituencies give donors both input and output legitimacy through facilitation of access to local engagement and support in the constituency, and recognition and confirmation of the effectiveness of the donor’s efforts. Similarly, these groups can threaten ex post reputational costs if they are neglected in allocation (either in providing benefits or avoiding negative externalities) by naming and shaming donors that have overlooked their needs.⁵

Key to the arguments above is the assumption of a pro-poor donor actor. If a donor is unconcerned with addressing needy constituencies their allocation is unlikely to be swayed by the political empowerment of those constituencies. Sub-national allocation of egoistic donors is likely to follow their own interests if those interests have a geographic component, perhaps access to resources or to accompany a subnational geo-strategic objective (Bohnke & Zurcher, 2013; Findley, Powell, Strandow, & Tanner, 2011). Alternatively, egoistic donors without geographic preferences for aid allocation may simply leave it to recipient country governments to decide where to allocate aid (Dreher, Langlotz, & Marchesi, 2017). Our argument is that it is the combination and relative strength of these three influences – donor motivation, recipient government political economy, and needy constituency empowerment – that can explain subnational aid allocation patterns. That said, when needy constituencies are unempowered, or donor’s interests are purely egoistic, the dance reverts to two players.

2.1. Subnational allocation of education aid in India: follow the caste

In order to evaluate our theoretical claims, we examine the case of World Bank education aid allocation in India. The donor, recipient and sectoral foci allow us to more precisely test the mechanisms we’ve outlined above. Focusing on the World Bank, as a multilateral donor actor, allows us to proceed with a reasonable assumption of donor pro-poor motivation. Bilateral and multilateral donors have been differentiated based on the principal-agent considerations the latter (Eichenauer & Hug, 2018; Milner & Tingley, 2013). While some argue that multilateral donors are captured by their powerful principals (Dreher & Jensen, 2007; Schneider & Tobin, 2013; Stubbs, Kentikelenis, & King, 2016), Copelovitch (2010) presents a theoretical argument and empirical evidence that multilateral donors both fall under the control of their powerful principals but also have areas of agency slack, although multilateral bureaucrats still may use that slack to please their main principal (Clark & Dolan, 2020). Despite these nuances, it is conventionally held that multilateral donors are more likely to be pro-poor in their decision-making (Clist, Isopi, & Morrisey, 2012; Thiele, Nunnenkamp, & Dreher, 2007).

From a recipient standpoint, focusing on India allows us to utilise the unique and persistent social structures of that country in determining and differentiating needy constituencies. The prevailing cause of social inequality in Indian society is caste identity⁶ determined by birth. Indeed, 45.9 per cent of scheduled tribes and 26.6 per cent of scheduled castes are in the lowest wealth bracket.⁷ Yet, SC/ST is not directly synonymous with wealth, and as such the classifications allows us to identify and compare empowered and non-empowered needy constituencies within the more general ‘poor’.

Finally, examining education aid enables us to concentrate on a distinctly ‘pro-poor’ sector (Thiele et al., 2007). Unlike other development sectors, there is a decent consensus that education aid ‘works’ at least when considering quantity metrics like enrolment or repetition rates (D’Aiglepierre & Wagner, 2013). While there is some evidence that in recent years education aid has become more entwined donor geo-strategic interests (Novelli, 2010), it remains a sector that is less likely to be driven by donors’ egoistic interests, especially given that payoffs from investment in education take years if not decades to materialise (Rao & Vadlamannati, 2011). Accordingly, focusing on World Bank education aid in India gives us a most likely case for observing the impact of empowered needy constituencies on patterns of aid allocation.

2.2. Caste, education, and the World Bank’s action in India

Substantial work has examined the impact of the caste system on access to education and education development (Borooah & Iyer, 2007; Halim, Yount, & Cunningham, 2016; Scaria, 2014; Tilak, 1979). This scholarship indicates that the caste system causes durable inequality in education and different levels of educational achievement, employment outcomes and access to economic resources (Borooah, 2012; Pellissery, Pampackal, & Bopaiah, 2015; Scaria, 2014) These inequalities persist despite provisions in the Indian Constitution to explicitly deal with this discrimination which emphasising that:

No citizens shall, on grounds only of religion, race, caste, sex, descent, place of birth, residence or any of them, be ineligible for, or discriminated against in respect of, any employment or office under the State.⁸

“Untouchability” is abolished and its practice in any form is forbidden. The enforcement of any disability arising out of “Untouchability” shall be an offence punishable in accordance with law.⁹

Despite these provisions, exclusion in education is still a main ritual marker of lower caste status (World Bank, 2011). A report from the Social and Rural Research Institute (2014) indicated that, as of 2014, 3.24 per cent of SC children and 4.20 per cent of ST children are still not enrolled in school.

The World Bank has a long history of education programs in India.¹⁰ Before the 1990s, the school system in India was mainly domestically financed. However, in the early 1990s, a widening gap between public expenditure and revenues required the Government of India to reduce expenditure on education (Tilak, 2008). Following the aid commitments from the World Conference on Education for All in 1990, international donors like the World Bank and the IMF increased their attention to basic education. Meanwhile, as a result of constitutional reform, school management responsibilities were gradually transferred to local bodies at the district, village and block levels (Colclough & De, 2010). Guided by the Eighth Plan Document (1992–1997), education initiatives became more targeted by shifting in focus from educationally backward states to districts.¹¹

Ensuring ‘the right to education’ has long been a goal of the World Bank (Oestreich, 2004). To protect ‘the right to education’, the World Bank has emphasised the rights to education in vulnerable groups in society, which includes women, refugees, those subject to involuntary resettlement, as well as indigenous people. Protection of these groups have been built into projects (Omprasad, 2016). Accordingly, the World Bank has paid strong rhetorical homage to the rights to education among socio-economic minorities when initiating and implementing education projects. In a report titled Learning to Be published in 1972, the World Bank argued:

The universal right to education – in which contemporary civilization takes such premature pride – is often refused, by a complete reversal of justice, to the most underprivileged. They are the first to be denied their right in poor societies, the only once deprived in the rich (Faure et al., 1972).

In the context of India, documents from the World Bank frequently acknowledge that the caste structure produces significant gaps in performance, with the high castes learning more and working more productively than the low castes (Hoff, 2016; Hoff & Pandey, 2011). Then World Bank president, Jim Yong Kim, noted in addressing the Vibrant Gujarat Summit that caste bias is always a concern of the World Bank and allocation of funds for entrepreneurs from among the scheduled castes should be ethically just and economically sound.¹² Similarly, expanded human development lending in the World Bank’s 2004 Country Strategy for India was predicated on the continuing substantial disparity of opportunity, particularly in the education, health and economic prospects of women and other vulnerable groups like SC and ST populations. Likewise, the Scheduled Caste and Scheduled Tribe Development Plan outlines mitigation measures to solve problems of exclusion in education. Thus, the official rhetoric surrounding the allocation of World Bank education aid in India contains an explicit and continued pro-poor focus, particularly for SC/STs.

However, paying homage to SC/STs in official documents may simply be ‘cheap talk’ that does not necessarily translate to observed patterns of aid allocation (Dreher et al., 2017). To understand if the World Bank’s apparent altruism translates to increased SC/ST allocation targeting, we need to understand the interplay between the Bank, the Indian government authorities, and the SC/ST groups themselves. The World Bank’s elementary education aid in India is delivered through projects. After investigating relevant documents of the World Bank’s elementary projects in India, it is evident that the World Bank signed project agreements with the Government of India to propose policy implementation in specific districts.¹³ DPEP financing was targeted to districts with female literacy rates below the national average of 39 percentage, as well as to districts where Total Literacy Campaigns have generated substantial enrolment increases in primary education.¹⁴ While all funded districts met one of these criteria, not all districts that met one of these criteria received DPEP funding.

At the level of implementation, the relationship between the World Bank and the Central Government of India is fixed by project agreements, which guarantees information flow and makes monitoring and accountability easier (Radelet, 2006). The implementation agencies of the Bank’s education projects in India are mainly local governments.¹⁵ Non-governmental actors also sought to influence the nature and implementation of these projects. The International Dalit Solidarity Network (IDSN) recommended governments take appropriate measures to ensure Dalits’ rights to equal participation and non-discrimination in education.¹⁶

Given the discussion above, we develop two expectations about pattern of the World Bank’s education aid allocation in India. First, given the explicit focus of both the World Bank and the Indian government on SC/STs, we expect increased allocation to those areas with high proportions of these marginalised poor above and beyond targeting towards the general ‘poor’. Second, given the mechanism of reducing search costs and increasing input and output legitimacy, we expect this effect to be amplified in locations where the SC/ST is politically empowered.

3. Data and methods

To examine our marginalised poor and political empowerment hypotheses, we apply cross-sectional data covering 593 districts from 29 states and seven Union Territories in India (Appendix 1). We use the World Bank Geocoded Aid Data v1.4.2 from AidData (2017) and the Census of India 2001 handbook sourced from the Ministry of Home Affairs, Government of India, that provides data on social and economic indicators at the district level. We formulate cross-sectional data by matching these two datasets and then adding additional control variables sourced from the Reserve Bank of India’s Database on Indian Economy. Since some of the data are not available for all districts and for all years, our dataset is unbalanced. We thus estimate:

(1) $W{B}_{di}={\beta }_{1}SCM{P}_{di}+{\beta }_{2}S{C}_{di}+{\beta }_3{X}_{di}+{\varphi }_i+{\epsilon }_{di}$(1)

(2) $W{B}_{di}={\beta }_{1}STM{P}_{di}+{\beta }_{2}S{T}_{di}+{\beta }_3{X}_{di}+{\varphi }_i+{\epsilon }_{di}$(2)

where in Equations (1)(1) $W{B}_{di}={\beta }_{1}SCM{P}_{di}+{\beta }_{2}S{C}_{di}+{\beta }_3{X}_{di}+{\varphi }_i+{\epsilon }_{di}$(1) and (2(2) $W{B}_{di}={\beta }_{1}STM{P}_{di}+{\beta }_{2}S{T}_{di}+{\beta }_3{X}_{di}+{\varphi }_i+{\epsilon }_{di}$(2) ), $W{B}_{di}$is the outcome variable of interest, $X_{di}$are the control variables (discussed below),${\varphi }_i$ is state fixed effects and ${\epsilon }_{di}$is error term. The $W{B}_{di}$is a dummy measure taking the value 1 if district d in state i has the World Bank elementary education aid allocation and 0 otherwise. We focus on one broad education initiative from the World Bank, the District Primary Education Project (DPEP) which was disbursed in seven different project waves from 1994 to 2001. Note that DPEP data is available at the project-level and doesn’t vary by year and hence our data is cross-sectional. While the AidData database has information on these waves, the geographic precision for many of these projects in that data is only at the state level. However, Azam and Saing (2017) identify which districts within states have received DPEP and we combine that with the AidData project records to create a measure of district-level aid.¹⁷ In total, 268 districts received at least one wave of DPEP financing in amounts ranging from roughly US$3 million to US$10 million when divided equally across districts. However, since the data does indicate aid amounts by individual districts, we use a dummy variable which denotes whether the district received World Bank aid or otherwise.¹⁸ While it would be of interest to test our hypothesis with other types of aid, this is the only large-scale project that we could reliably identify at the district level.

Since the dependent variable is a dummy measure an ideal estimator to utilise would a non-linear maximum likelihood estimator. Using a non-linear maximum likelihood estimator does not allow us to include state-fixed effects mainly due to the well-known incidental parameter problem (Lancaster, 2000). To circumvent this problem, we follow Eichengreen and Leblang (2008) and estimate a linear probability model, which allows us to control for state-fixed effects. All our models, thus, are estimated using a linear estimator which controls state-fixed effects. In robustness tests, we also estimate a probit model without state-fixed effects.

The two key variables of interest in Equations (1)(1) $W{B}_{di}={\beta }_{1}SCM{P}_{di}+{\beta }_{2}S{C}_{di}+{\beta }_3{X}_{di}+{\varphi }_i+{\epsilon }_{di}$(1) and (2(2) $W{B}_{di}={\beta }_{1}STM{P}_{di}+{\beta }_{2}S{T}_{di}+{\beta }_3{X}_{di}+{\varphi }_i+{\epsilon }_{di}$(2) ) are the SC and the ST population shares in district d, state i respectively. The data for both SC and ST and total population for each district is sourced from the Government of India 2001 census handbook. The SC and ST MPs is a count of number of elected MPs from SC and ST reserved constituencies and 0 for non-reserved constituencies. The challenge we encounter in compiling this data is that the electoral constituencies in India do not overlap with administrative districts’ boundaries in the states, a problem also faced by other studies in the literature (Gehring, Kauffeldt, & Valamannati, 2019). We make use of the documents available at the Election Commission of India which provide information about the boundaries of administrative districts and electoral constituencies in each state which was in turn used to match the individual constituencies reserved for SCs and STs to the administrative districts. $\mathrm{l}\mathrm{n}{\left(\mathrm{W}\mathrm{B}\right)}_{\mathrm{d}\mathrm{i}\mathrm{t}}={\beta }_1{\left(\mathrm{S}\mathrm{C}/\mathrm{S}\mathrm{T}\times \mathrm{S}\mathrm{C}/\mathrm{S}\mathrm{T}\mathrm{M}\mathrm{P}\right)}_{\mathrm{d}\mathrm{i}\mathrm{t}}+{\beta }_2\mathrm{S}\mathrm{C}/\mathrm{S}{\mathrm{T}}_{\mathrm{d}\mathrm{i}\mathrm{t}}+{\beta }_3\mathrm{S}\mathrm{C}/\mathrm{S}\mathrm{T}\mathrm{M}{\mathrm{P}}_{\mathrm{d}\mathrm{i}\mathrm{t}}+{\beta }_4\mathrm{S}\mathrm{C}/\mathrm{S}\mathrm{T}\mathrm{c}\mathrm{e}\mathrm{n}\mathrm{s}\mathrm{u}{\mathrm{s}}_{\mathrm{d}\mathrm{i}\mathrm{t}}+{\beta }_5{\varphi }_{\mathrm{d}\mathrm{i}\mathrm{t}}+{\lambda }_{\mathrm{t}}+{\delta }_{\mathrm{i}}+{\mathrm{\partial }}_{\mathrm{d}}+{\epsilon }_{\mathrm{d}\mathrm{i}\mathrm{t}}\dots \left(5\right)$

Next, we further operationalise our political empowerment hypothesis by considering the differential effect of SC and ST MPs when they are in government. To do this, we estimate a model in line with Vadlamannati (2015) and Khemani (2007).

(3) $W{B}_{di}={\beta }_1{\left(SCMPs\times Affiliation\right)}_{di}+{\beta }_{2}SCMP{s}_{di}+{\beta }_{3}Affiliatio{n}_{di}+{\beta }_4{X}_{di}+{\varphi }_i+{\epsilon }_{di}$(3)

(4) $W{B}_{di}={\beta }_1{\left(STMPs\times Affiliation\right)}_{di}+{\beta }_{2}STMP{s}_{di}+{\beta }_{3}Affiliatio{n}_{di}+{\beta }_4{X}_{di}+{\varphi }_i+{\epsilon }_{di}$(4)

where $Affiliatio{n}_{di}$ is an indicator of political affiliation that equals 1 when the political affiliation of SC and ST MPs from district d belongs to the same political party as that governing at the centre, and 0 otherwise. Note that we control for population share of SCs and STs in Equations (3)(3) $W{B}_{di}={\beta }_1{\left(SCMPs\times Affiliation\right)}_{di}+{\beta }_{2}SCMP{s}_{di}+{\beta }_{3}Affiliatio{n}_{di}+{\beta }_4{X}_{di}+{\varphi }_i+{\epsilon }_{di}$(3) and (4(4) $W{B}_{di}={\beta }_1{\left(STMPs\times Affiliation\right)}_{di}+{\beta }_{2}STMP{s}_{di}+{\beta }_{3}Affiliatio{n}_{di}+{\beta }_4{X}_{di}+{\varphi }_i+{\epsilon }_{di}$(4) ) respectively. If political empowerment stems from SC and ST MPs being politically aligned with the central government, we could have β₁ > 0 and β₂ not significantly different from 0. Conversely, if partisan identity of the SC and ST MPs does not matter for the ruling party in centre, then β₁ would be indistinguishable from 0. Once again, we employ linear probability controlling for state-fixed effects estimators using Huber-White corrected robust standard errors.

Finally, the vector $X_{di}$includes control variables at the district level in India which are gleaned from the literature on aid allocation at the subnational level (Briggs, 2017; Dreher, Nunnenkamp, & Thiele, 2008; Nunnenkamp et al., 2016). In selecting the controls, we try to avoid the ‘garbage can’ approach (Achen, 2005) and follow a conservative strategy of accounting only for known factors that may confound the effect of SC and ST population share, such as level of income, spending on education and the literacy rate. Accordingly, we include district-wise population (log) as larger districts might need more resources to obtain visible effects of aid provision (Nunnenkamp et al., 2016). In considering general district-level need, we consider three proxies. First, we expect the World Bank to provide more education aid to districts with a larger rural population share where educational infrastructure is limited (Dollar & Levin, 2006). Next, we expect districts with lower levels of literacy to attract more aid to education from the World Bank, serving as a measure of educational need. Following Dreher et al. (2008), we include male and female literacy rate drawn from the 2001 Census handbook. Finally, we use satellite night-time lights images as a proxy for the economic development of the districts as there are no GDP estimates at the district level in India (Gehring et al., 2019). Henderson, Storeygard, and Weil (2012) show how to calculate night-time lights data and suggest that it is correlated with official GDP growth data. We use average visible, stable, light on cloud free nights, collected by the F16 satellite for the years 1992 and 2001. We then compute the log sum of lights using zonal statistics within each district to proxy for economic development. In the probit model robustness tests, we also control for state capacity variables which vary among states but not by district within each state. These include government’s total and primary education expenditure (log) at the state-level and the distance from each district to the capital of that state measured in kilometres (log) (Hoelscher, Miklian, & Vadlamannati, 2012; Nunnenkamp et al., 2016, Vadlamannati, 2011) capturing remoteness of districts. The descriptive statistics are provided in Appendix 2 and details on data definitions and sources in Appendix 3.

4. Empirical results

Table 1 reports our main results. Columns 1–2 present the results on the impact of SC and ST population shares on the probability of World Bank educational aid allocation. Columns 3–4 examine whether the probability of World Bank aid allocation in Indian districts is conditional on the presence of SC and ST MPs, controlling for their population shares. As seen from column 1, our parsimonious model examining the unconditional effects of SC and ST population shares, the SC population share is positive and significantly different from zero at the 1 per cent level. The substantive effect suggests that, holding all other control variables constant at their mean values, a mean plus standard deviation increase in SC population share is associated with a 29 per cent increase in the likelihood of World Bank aid allocation, which is significantly different from zero at the 1 per cent level. These results are robust to the inclusion of control variables in column 2. Once again, the substantive effect suggests that a mean plus standard deviation increase in SC population share is associated with a 21 per cent increase in the likelihood of World Bank aid allocation, while increasing the SC population share to its maximum value (50.11) increases the chance of securing World Bank aid by 45 per cent.

Table 1. World Bank education aid allocation, SC/ST population share and SC/ST MPs

	(1)	(2)	(3)	(4)
SC Population share (Current)	0.0118***	0.00867**	0.0104***	0.00728**
	(0.00321)	(0.00338)	(0.00346)	(0.00356)
ST Population share (Current)	0.00417***	0.00159	0.00311**	0.000797
	(0.00113)	(0.00125)	(0.00123)	(0.00135)
SC MPs			0.0491	0.0466
			(0.0438)	(0.0437)
ST MPs			0.0994*	0.0626
			(0.0558)	(0.0555)
Population (log)		0.0523		0.0430
		(0.0350)		(0.0356)
Rural Population share		0.122		0.126
		(0.166)		(0.166)
Literacy Rate		−0.0125***		−0.0123***
		(0.00206)		(0.00208)
Night Light (log)		−0.00430		−0.00423
		(0.00339)		(0.00340)
Constant	0.394*	0.480	0.412*	0.627
	(0.222)	(0.646)	(0.221)	(0.650)
R-squared	0.361	0.404	0.365	0.406
Total Observations	597	545	597	574
Estimator	OLS-FE	OLS-FE	OLS-FE	OLS-FE
State Fixed Effects	Yes	Yes	Yes	Yes
Number of States	29	29	29	29
Number of Districts	597	545	597	574

Notes: Robust standard errors in parenthesis; Statistical significance: ***p < 0.01, **p < 0.05, *p < 0.1.

Next, we also find a positive and statistically significant relationship between ST population share and likelihood of World Bank aid allocation. However, when we control for other variables in column 2, the effect is no longer statistically significant. There could be multiple reasons for this finding. One is that there are fewer districts with substantial ST populations and there may simply be insufficient variation to detect an effect upon adding controls. A second plausible explanation is that, compared to SC population, the majority of the ST population resides in rural, and often remote, areas (Chin & Prakash, 2011). The poverty rate of this community is twice that of SC population which tends to inhabit urban areas. As such, we might expect that World Bank exerts efforts to reach out to these communities by focusing on rural areas to improve ST well-being. However, the reach of the Indian government machinery to some of these rural areas is limited. In a UN survey report, Sujatha (2000) points out the difficulties faced by the government administrative apparatus to take some of the government educational schemes to the ST districts which are in remote parts of the country, findings echoed by Nair (2007).

Next, we examine whether the World Bank aid allocation in Indian districts is conditional on the presence of SC and ST MPs, controlling for their population shares. In column 3, we introduce new variables, that is, if a district is represented by a SC or ST MP without any control variables. As seen there, we find positive effects, but these are only statistically significant at the 10 per cent level for ST MPs. These findings are in line with Chin and Prakash (2011), and Pande (2003), who find that, at the state level, political reservations for minority groups, especially STs in India, does reduce overall poverty.¹⁹ The substantive effects suggest that a district represented by ST MPs is associated with roughly 10 per cent increase in chance of securing World Bank aid allocation. However, the fixed effects model with control variables in column 4 soaks up the statistical significance of ST MPs. Notice that the SC population share continues to remain positive and significantly different from zero at the 1 per cent and 5 per cent level respectively across the models in Table 1.

With respect to control variables, we do not find population, nigh-time light or rural population share to be significant correlates of securing development aid projects in education sector from the World Bank. However, we find that the literacy rate in districts is strongly associated with World Bank aid allocation, in line with Nunnenkamp et al. (2016). Our results suggest that a mean plus standard deviation increase in the literacy rate is associated with a 96 per cent increase in the likelihood of World Bank aid allocation, which is significantly different from zero at the 1 per cent level in both column 1 and 4. The effects are substantially large as the DPEP explicitly targeted this metric. The positive findings on SC and, to a lesser extent, ST measures when accounting for the literacy rate suggests that these allocation patterns are above and beyond patterns based solely on educational need.

4.1. Political agency effects

To test the political agency hypothesis, we present the results on political affiliation of SC and ST MPs in Table 2. In columns 1 and 2 we present the results on SC and ST MPs political affiliation with and without control variables. As seen in column 1, the interaction results suggest that an SC MP political affiliation with the same political party as the central government increases the probability of securing World Bank aid allocation. For instance, districts represented by SC MPs see an increase in likelihood of World Bank aid allocation by 14 per cent when they are affiliated to the ruling party in the central government. It is noteworthy that the interaction effect of ST MPs with political affiliation is statistically insignificant. One plausible explanation, apart from the other explanation of remoteness discussed previously, could be that number of ST MPs affiliated with ruling party in central government are lower compared to their SC counterparts. For instance, compared to 37 per cent of SC MPs who are not affiliated to ruling party in central government, about 52 per cent of ST MPs were not part of the government in centre during this period.

Table 2. Partisanship and World Bank education aid allocation

	(1)	(2)
SC MPs Х Political affiliation	0.138*	0.167**
	(0.0742)	(0.0783)
SC MPs	−0.0427	−0.0620
	(0.0531)	(0.0583)
SC Population share (Current)	0.0113***	0.00793**
	(0.00353)	(0.00361)
ST MPs Х Political affiliation	0.165	0.180
	(0.121)	(0.135)
ST MPs	−0.0233	−0.0755
	(0.112)	(0.129)
ST Population share (Current)	0.00333***	0.00101
	(0.00122)	(0.00134)
Political Affiliation	−0.0562	−0.0632
	(0.0585)	(0.0604)
Population (log)		0.0438
		(0.0361)
Rural Population share		0.141
		(0.168)
Literacy Rate		−0.0124***
		(0.00206)
Night Light (log)		−0.00439
		(0.00355)
Constant	0.433**	0.631
	(0.217)	(0.651)
R-squared	0.37	0.413
Total Observations	597	574
Estimator	OLS-FE	OLS-FE
State Fixed Effects	Yes	Yes
Number of States	29	29
Number of Districts	597	574

Notes: Robust standard errors in parenthesis; Statistical significance: ***p < 0.01, **p < 0.05, *p < 0.1

The other plausible explanation could be technical. It is noteworthy that both interaction variables are strongly correlated with their respective constituent terms namely, SC and ST MPs.²⁰ We follow Roodman’s (2008) approach to address the multicollinearity problem. We estimate the interaction specifications as reported in Table 2 without their respective constituent terms (SC and ST MPs) we find both interaction terms to be positive and statistically significant at the 5 per cent and 10 per cent levels, respectively (results in Appendix Table H). This suggests that multicollinearity could be driving the insignificant results on the ST interaction variable.

Notice that both SC and ST population shares exert positive and significant effect on the probability of securing World Bank aid. These results are consistent with the results reported in column 1 in Table 1 in which control variables are not included. These results on interaction effects remain robust in column 2 when we estimate interaction effects controlling for other control variables. As seen there, once again the positive significant effect of SC MPs conditional upon political affiliation to centre on World Bank aid allocation remain robust. These effects are significantly different from zero at the 5 per cent level, while the interaction results on ST MPs continue to remain statistically insignificant. Interestingly, the constituent terms of interactions on their own remain statistically insignificant. This is in line with our previous findings that SC and ST MPs on their own (that is, when political affiliation variable is set equal to 0) have no significant effect in explaining World Bank aid allocation.

Overall, our results suggest that it is the SC MPs political relationship with the centre that accounts for significant variation in the probability of aid allocation of World Bank. Notice that the interaction results on SC MPs, in particular, are net of other important control variables suggesting that political representation matters when accounting for economic or educational needs. The control variables show the same effects as those reported in Table 1. These results give further credence to our political empowerment hypothesis. Overall, we take these allocation patterns as evidence that the World Bank has sufficient control to ensure aid programs, broadly, go to the ‘needy’ areas. However, within those parameters, we find evidence of constituency agency – it is politically empowered ‘needy’ areas that get the funding.

4.2. Robustness checks

We subject our main findings to several robustness checks, full discussion, and tables of which can be found in the supplemental online appendix. There, we account for possible endogeneity between SC/ST population shares and reserved SC/ST parliamentary seats, use a probit estimator, use financing amounts as the outcome variable, control for outlier districts in terms of number of SC/ST parliamentarians, and use alternative clustering for our standard errors. Our main results are robust to all these alternative approaches.

5. Conclusion and discussion

This paper sheds further light on the politics of subnational aid allocation. Notably, we have shifted the focus from the supply side of aid allocation to the dynamics of recipient demand. We have argued that marginalised constituencies, above and beyond the general poor, may be able to influence aid allocation, especially if they are politically empowered. Our findings give varying levels of support to these contentions. While there is some evidence that areas with high proportions of marginalised constituencies – in our case scheduled castes and tribes in India – receive World Bank education aid above and beyond what we might expect if allocation was based on general measures of need, there is strong and robust support that when these groups are politically empowered, the chances of receiving aid increases significantly. Notably, if an Indian district is represented by a SC member of parliament who is aligned with the governing party in the centre, the likelihood of receiving aid from the World Bank increases by 17 per cent compared to districts not represented by an SC MP aligned with the government. We stress, however, that our findings show patterns of allocation, not causal relationships. Work which uses data which allows for casual inference techniques, and/or qualitative data which sheds light on our proposed causal pathways, would be a valuable step forward.

Our findings have two implications for the broader literature of aid allocation. First, in line with theories of multilateral donors with sufficient ‘agency slack’, we find that the World Bank may be able to match action to rhetoric and target its assistance to localities with disadvantaged populations. Second, our findings also suggest that the agency of the marginalised population may influence allocation behaviour. Otherwise marginalised constituencies that can ‘speak for themselves’, through representation in government, can influence aid allocation behaviour. This might happen for several reasons, including the ability to reduce search and transaction costs, facilitate donor access, and/or provide donors with input and output legitimacy. These findings broaden our understanding of how international, multilateral, donors interact with both national and subnational actors in recipient countries. However, they also sound a cautionary note. If sub-national aid allocation depends on the (relative) empowerment of needy communities then the ‘squeaky wheels’ may ‘get the grease’ while the most marginalised continue to be overlooked.

Acknowledgements

We thank Editor and two anonymous referees for helpful comments and suggestions. We also thank Indra de Soysa, Christopher Kilby, Nishith Prakash, Jonathon Moses, Hannes Öhler, Artur Tamazian and participants at the Sustainability & Development Conference in Ann Arbour, Michigan 2019 for helpful comments.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Supplementary materials

Supplementary Materials are available for this article which can be accessed via the online version of this journal available at https://doi.org/10.1080/00220388.2020.1806247.

Notes

1. For our purposes political empowerment may mean access to formal institutional structures such as key parliamentary or administrative roles but may also simply be constituencies that are well-organised and empowered through non-governmental channels. We will elaborate on this in the empirical section below.

2. It is important to note that local, empowered, representatives may also take advantage of the information asymmetry to misrepresent their needs, as Adedokun (2015) argued occurred, at times, in the Haiti earthquake crisis. However, even when the need is misrepresented, this would still result in increased allocation to those areas.

3. https://www.worldbank.org/en/topic/communitydrivendevelopment accessed on 11 April 2020.

4. For example, in discussing a World Bank coffee project in Papua New Guinea, a local government official proclaimed, ‘We thank the World Bank … for coming down to the village level to save our coffee gardens.’ Further information can be found at < http://www.looppng.com/content/villagers-praise-world-bank-and-partners-coffee-rehab-work> Accessed on 10 January 2018.

5. As an example, a story of a local from the Tanzanian village of Mlanda who passes a barren pump on a 2 km walk to retrieve water, despite World Bank promises to improve the water supply. Further information can be found at <https://www.pri.org/stories/2014-11-24/world-banks-water-failure-tanzania> Accessed on 11 October 2018. More directly, a group of 103 families from three villages in the Sindhuli district of Nepal complained to Bank Management over negative externalities from a power-line transmission project which was ultimately picked up in several news sources. https://www.worldbank.org/en/news/press-release/2015/07/13/board-discusses-nepal-inspection-panel Accessed on 11 April 2020.

6. The caste system entails a division of labour, wherein Brahmin is the priestly class, Kshatriya is the military class, Vaishya is the merchant class and Shudra comprises artisans and menial workers. Outside this system falls ‘Dalits’ and indigenous inhabitants (Pellissery et al., 2015).

7. Government of India (2018).

8. The Government of India (1950).

9. The Government of India (1950).

10. While Nunnenkamp et al. (2016) found that the World Bank prefers districts where foreign direct investors may benefit from projects related to infrastructure, they did not explicitly examine the allocation of education aid.

11. Relevant documents about The National Plans from the Government of India can be accessed at < http://planningcommission.nic.in/plans/planrel/index.php?state=planbody.htm>.

12. The Times of India (2015).

13. There were 15 elementary education projects in India from the World Bank from 1991 to 2011. And each project offered series of documents including Memorandum & Recommendation of the President, project agreements, credit agreements, implementation completion and results reports and implementation report reviews. The authors reviewed these documents and found that project agreements are normally singed between the International Development Association and Government of India.

14. World Bank (1994), District Primary Education Project, available at: http://documents.worldbank.org/curated/en/601621468771640834/pdf/multi0page.pdf Accessed on 2 February 2019.

15. We explored the World Bank education projects documents in India from 1992 through 2014 to find that implementation agencies in these projects were all state governments of India.

16. https://idsn.org/key-issues/education/accessed on 11 April 2020.

17. While the DPEP project is somewhat dated, this district-level coding allows for the subnational analysis we conduct here which would not be feasible with later projects that only have information available at the state level. As our main argument is that it is relatively time invariant, recipient-side, socio-political dynamics that are influencing ultimate allocation locations, we think any insights gleaned from our analysis would still be relevant in more contemporary settings.

18. The common practise is to divide the project aid amount evenly across the locations. We do this as a robustness check.

19. Of note from Pande (2003) is that areas with SC/ST legislator reservations received less education funding from the Indian government. This might suggest that the World Bank program simply filled in for the Indian Government and/or the Indian Government reduced its spending in response to the SC/ST targeting of the World Bank.

20. The correlation of interaction term${\left(SCMPs\times Affiliation\right)}_{di}$ with SC MPs is 0.80, while the ${\left(STMPs\times Affiliation\right)}_{di}$ is correlated with ST MPs at 0.81.