Skip to main content
Open Access

Do Citizens Enforce Accountability for Public Goods Provision? Evidence from India’s Rural Roads Program


This article investigates voter responsiveness to the world’s largest rural roads program, a highly visible development program that improved connectivity for one-third of humanity that previously lacked road access. Investigating 180,000 roads provided across half a million Indian villages aggregated across multiple elections over the last 20 years, the article finds that road provision fails to boost electoral support for the ruling party. Exploiting population-based implementation rules that partly determine road allocation, instrumental variable regressions show that voters remain unresponsive to exogenous road provision. Exploiting subnational variation in implementation and political alignment, analysis shows that factors that breakdown the accountability chain, such as quality, salience, myopia, corruption, or attribution concerns, do not explain these results. The findings suggest that weak accountability presents a more enduring challenge to democracy than assumed in theoretical models and policy interventions.

Citizens in low- and middle-income countries overwhelmingly report the provision of public goods and services such as roads, water, and electricity as their top policy issue (Grossman and Slough 2022). Yet it is unclear whether citizens use elections to enforce accountability for service delivery by punishing incumbents who underprovide these crucial services while rewarding those who improve access.1 Studies that investigate electoral accountability focus on cash transfers or financial grants that are targeted at individuals (Blattman, Emeriau, and Fiala 2018; Manacorda, Miguel, and Vigorito 2011; Pop-Eleches and Pop-Eleches 2012), despite the fact that politicians often play no role or lack discretion to enable such transfers (Imai, King, and Rivera 2020). Other scholars have investigated policies that are lower priorities for voters (Boas, Hidalgo, and Toral 2021) or have studied citizens’ responses to audits that have generated unique media sensations (Ferraz and Finan 2008; Larreguy, Marshall, and Snyder 2020). Even if voters respond to policies such as cash transfers or audits, this does not necessarily mean that they will be similarly receptive to signals of everyday provision of public goods and services that often do not make sensational headlines and are provided at the community level.

This article conducts the largest investigation of voters’ electoral responsiveness to public goods provision at an unprecedented time, population, and geographic scale using the case of India’s $41 billion rural roads program, the Pradhan Mantri Gram Sadak Yojna (Prime Minister’s Rural Roads Program, or PMGSY). The PMGSY improved connectivity for one-third of humankind by constructing close to 180,000 rural roads, making it the largest visible change in India’s rural corridors. PMGSY is also a strong case for observing electoral accountability in a developing context. The scale of road construction alone makes it impossible to ignore. An average Indian village has a diameter of 2.1 kilometers, and the average road built through this program is 4.4 kilometers long (Adukia, Asher, and Novosad 2020). Additionally, this highly visible development program included accountability and transparency features to improve widespread access to high-quality program information. Like the provision of most other public goods and development schemes in India, subnational state governments were responsible for implementing PMGSY. PMGSY formally included events where politicians could claim credit for providing roads, and evidence confirms political influence in road allocation (Bohlken 2021). At the same time, rich qualitative evidence from a variety of sources underscores voters’ awareness of and their political attribution of this program to state governments. Furthermore, existing studies exploit exogenous variation in road provision to provide causal evidence that voters have benefited economically and socially from PMGSY roads (Adukia et al. 2020; Aggarwal 2018, 2021; Asher and Novosad 2020). This article enables knowledge accumulation in this research agenda by investigating voters’ electoral response to the PMGSY and contributes to the study of large-scale infrastructure programs more broadly.

Studying electoral accountability for providing basic services is important for two reasons. First, multi-billion-dollar programs to provide basic services have proliferated in the developing world (World Bank 2004). However, lacking electoral incentives, politicians are more likely to abandon these programs or invest their effort in high-visibility tasks that may not necessarily align with citizen welfare or priorities (Harding and Stasavage 2014). Second, governments have invested sparse state resources in transparency and accountability initiatives, which are increasingly built into development programs to improve citizen accountability (Gaventa and McGee 2013). So, it is important to ask whether these costly, state-led efforts are getting the job done.

Theoretically, there is a higher likelihood of observing accountability for road provision, relative to more complex services like education or health care (Harding 2015; Mani and Mukand 2007). However, recent research on information and accountability in developing countries yields opposing conclusions as to whether citizens will hold politicians accountable for providing access to crucial services even when conditions are favorable (Dunning et al. 2019). One strand of research raises the concerns that the current theoretical models should be revised (Adida et al. 2017; Martin and Raffler 2021; Weaver 2021), while the other suggests that there are breaks in the information chain that make performance-based voting harder but does not question the current theoretical paradigm (Adida et al. 2020; Boas et al. 2021).

The PMGSY, by providing roads at an unprecedented scale, provides a unique opportunity to advance the research on accountability. After discussing in detail how the PMGSY offers a compelling theoretical opportunity of observing accountability for service delivery in India, this article investigates the electoral response to this program. To do so, I leveraged data on 180,000 all-weather rural roads in India built between 2001 and 2017, which I have merged with constituency-level data on national and state elections. This data set covers more than 90% of India’s population and a significant duration of the road program. Crucially, rich data on road quality and cost enable me to investigate whether corruption moderates voter response, which is a key explanation behind weak accountability and incumbency disadvantage (Klašnja 2015). I begin by estimating first-difference (FD) ordinary least squares (OLS) regressions. To improve inference, I analyze the change in the incumbent party’s vote share at the state constituency level, within the district-administration level at which the roads plan are prepared and construction contracts are sanctioned. Because voters may reward program launch, I also investigate voter response in national elections.

Using FD OLS regressions, this article finds that road provision does not boost electoral support for the ruling party. This null result holds across India, at various electoral levels, and across different time periods. As first deployed in Aggarwal (2018, 2021) and Asher and Novosad (2020), I exploit the population-based implementation rules that partly determine road allocation, to identify the effects of road provision at the constituency level using an instrumental variables (IV) regression. These existing studies provide evidence for the robustness of using population thresholds for the purpose of identifying the effects of road provision at the village and district level. I exploit the same design to construct an eligibility instrument that determines road provision at the constituency level. Data lend strong support for the exclusion restriction. IV results reinforce the null findings from the OLS analysis, suggesting that the results are robust to endogeneity concerns. Results from the village-level analysis in India’s most poorly connected and populous state, Uttar Pradesh (UP), confirm that constituency-level results do not average out over positive and negative spillover effects.

The main objective of this article is to provide a test of the sharp null hypothesis of real-world electoral accountability at scale. Additionally, I follow the literature to rule out key reasons for accountability failures that may explain why road provision does not have electoral effects. While investigating theoretical challenges to the political agency model is outside the scope of this article, I focus on explanations where there is breakdown within the accountability chain, such as lack of information, corruption, or poor attribution. Exploiting subnational variation in implementation and political alignment, I find that key explanations that center on quality or salience concerns, myopia, information- and media-rich environments, or a lack of clarity of responsibility are not the roadblocks that lead to failures in the accountability chain. Road provision, despite being visible and a top priority for voters, is unlikely to be a factor in their voting decision.

This study contributes to the literature on the political economy of accountability in several ways. First, the findings improve our understanding of the scale of the accountability problem in developing countries. Surprisingly few studies investigate the electoral effects of large-scale public-goods programs that also are among voters’ priorities (Grossman and Slough 2022). Second, an advantage of this aggregate analysis is the ability to use large-scale administrative data and real-world voting outcomes to test a sharp null hypothesis, complementing ongoing field experimental research. Field experiments require fewer identifying and ecological assumptions and are better suited to investigate mechanisms that require individual-level data, such as voters’ ethnic or partisan identities.2 However, they have a shorter time and geographic span, and a majority rely on self-reported survey data to measure voting intentions since real-world voting behavior is never directly observed (Dunning et al. 2019). This study bolsters support for the null finding from this agenda by showing that a local infrastructure provision program deploying transparency initiatives—arguably a highly powered information treatment—remains insufficient for increasing accountability. Third, this article contributes to the study of the multi-billion-dollar infrastructure development programs that have proliferated in developing countries. Research investigating these programs shows that state coordination failures lead to half-finished and abandoned projects (Williams 2017). This article offers an explanation as to why politicians ignore these failures: voter unresponsiveness discourages the type of costly political oversight that improves program implementation (Gulzar and Pasquale 2017; Raffler 2022). As they stand, the findings bolster concerns about the prospects for electoral accountability and suggest that in-built transparency initiatives common to development programs worldwide remain insufficient for increasing accountability.

Accountability and Public Goods Provision

A key goal of democracy is to ensure that all people have access to the basic goods and services necessary to live a dignified life. Despite decades of economic growth and democratically elected governments, many developing countries have failed to provide the majority of their citizens with equitable and high-quality access to basic necessities such as water, roads, and electricity (Pande 2020). Perhaps unsurprisingly, the provision of basic public goods and services also consistently ranks as a top concern for citizens. Reviewing cross-national survey data, Grossman and Slough (2022, 133) find that “sizable shares—in some cases even a majority—of the population in many countries cite [public goods and services as their top] issues.”

The chronic lack of basic services in the face of high voter demand raises concerns about the health of democratic accountability. Yet, despite these pressing concerns, the question of whether voters hold politicians accountable for providing local public goods and basic services in the real world remains understudied.3 Studying this question is urgent because a lack of voter responsiveness to welfare programs can divert politicians’ attention to high-visibility tasks that are not in line with voter welfare (Ashworth and Mesquita 2006; Harding and Stasavage 2014).

Theoretically, recent research on information and accountability has also complicated the political agency model’s straightforward expectation that voters should reward or punish politicians for providing or underproviding basic public services (Ashworth 2012). As per the standard political agency model, voters observe performance signals and then evaluate whether the politician is a good or bad type. Voters may rationally rely on shortcuts such as a candidate’s ethnic identity to inform their vote choice when they lack information on politicians’ competence and performance and candidates cannot commit to improving citizens’ well-being. Conversely, when voters receive clear performance signals, they are more likely to respond electorally; this feedback loop ensures that politicians remain attuned to voters’ needs. Building further on this insight, theoretical models conclude that highly visible and easy-to-evaluate public goods such as roads, electricity, or water provide clear performance signals and, therefore, are particularly favorable to accountability (Mani and Mukand 2007). Yet, a multisite and preregistered meta-analysis of the effects of voter information campaigns conducted in six vastly different countries concluded that the overall effect of providing standardized performance information is weak (Dunning et al. 2019).

Scholars have responded to these unexpected findings in two ways, yielding opposite conclusions regarding whether voters will enforce accountability for public goods provision when performance signals are clear. Scholars have also raised concerns about the premise of political agency models. Indeed, if citizens are unresponsive to new facts and reluctant to update their beliefs, it suggests that these traditional models are incorrect regarding citizens’ appetites for information and their propensity to use performance signals in exercising their voting choice. Adida et al. (2017) challenge the view that ethnicity is a heuristic that substitutes for a lack of information, demonstrating that ethnically motivated reasoning conditions how voters process new political information. Using rich data from quasi- and field experiments, Boas et al. (2021) find that voters punish high-quality education signals in Brazil because voters hold politicians accountable not only for their competence but also for their representation of potentially conflicting interests.

In addition, recent research highlights novel ways in which weak institutional context lowers the likelihood of observing accountability. A key explanation centers on the nature of challengers and the lack of incumbency advantage in developing countries. Voters’ strong preference for challengers regardless of incumbent quality, which is often a result of weak institutional accountability in nonelectoral institutions, can weaken the electoral accountability relationship. Weaver (2021) finds that mayors face a significant incumbency disadvantage in Peru and shows that neither good performance nor voters’ access to performance information enables mayors to overcome it. Voters prefer challengers regardless of incumbent quality, but it is voter trust in accountability institutions that attenuates incumbency disadvantage. Uppal (2009) also reports a strong and rising incumbency disadvantage in India but finds that this disadvantage is higher in states with lower provision of public goods. Klašnja (2015) suggests that corruption can hinder performance-based voting as voters prefer inexperienced challengers over experienced but corrupt incumbents. Eggers and Spirling (2013) argue the opposite: incumbency disadvantage can arise if narrowly elected incumbents are less appealing and lower quality than challengers on average. They argue that these scenarios are particularly likely in developing countries. If voters prefer challengers over incumbents for reasons such as corruption, a lack of access to basic services, or a lack of trust in state institutions that hold politicians accountable for their corrupt behavior, places where corruption is (less) rampant and services less (more) accessible should see a (lower) greater incumbency disadvantage.

Recent research has further called into question the veracity of the political agency model. Martin and Raffler (2021) highlight an additional dimension that moderates accountability: citizens allocating responsibility between politicians and bureaucrats. Using survey experiments in Uganda, they demonstrate how, when citizens believe that politicians have limited capacity to control bureaucrats, they are less likely to believe that government performance is a good signal of the incumbent’s quality. Frey (2022) raises new concerns about the accountability logic in Brazil’s clientelistic setting, where he finds that development programs, by permanently boosting voters’ incomes, reduced the voters’ likelihood to vote for incumbents. In summary, this body of research suggests we should revisit our expectations about any informational or other policy effort aimed at increasing electoral accountability, regardless of its execution, design, or source.

Yet, particularly in the light of null results (Dunning et al. 2019), research using field experiments to study accountability has also raised concerns about the design of commonly used informational interventions. Studies have argued that the complex chain of conditions linking information to accountability makes it impossible to determine whether voters do not process information or whether the way information is supplied challenges dissemination. In other words, this line of argument theorizes that, if execution challenges did not exist, we would observe voters acting to hold politicians accountable, thus confirming the theoretical predictions from the political agency model. Bhandari, Larreguy, and Marshall (2023) find that Senegalese voters care principally about such local outcomes as projects and transfers, rather than information on legislative efforts or attribution. They show that voters find temporally benchmarked local performance outcomes particularly informative, as these help voters to parse out common shocks that affect everyone as well as update about the absolute quality level of other politicians who resemble challengers. Boas et al. (2021) find Brazilian voters receptive to nonpartisan information about education quality provided in partnership with the state accountability institution. Adida et al. (2020) demonstrate that voters react strategically to information and that salience and coordination are necessary for information to influence voter behavior. In summary, this research illustrates how widely disseminated policy information concerning goods and services voters care about can enable accountability.

To summarize, the literature yields mixed expectations about whether voters enforce accountability for the provision of public goods and services. Political agency models suggest voters are likely to respond to the provision of public goods and services. Theoretically, local public goods provision, and roads in particular, is a favorable case for observing accountability. Harding (2015) shows that accountability for road provisions occurs even in contexts where patronage or ethnic politics remains pervasive. Voter responsiveness can align a politician’s incentive to remedy the lack or poor quality of a good or service. However, weak institutional context can remain an impermeable barrier, lowering the likelihood that voters respond electorally to performance signals. Investigating voter response to public service delivery can delimit the accountability problem more precisely and improve our understanding of why politicians implement or abandon development programs.

India’s Rural Roads Program

Before PMGSY, one-third of the world’s people who lacked access to an all-weather road lived in India. From 2000 to 2018, India built over 180,000 all-weather roads spanning over 550,000 kilometers at a cost of more than US$41 billion. Indian media has widely documented the program’s success and positive impact on rural citizens. Building on existing scholarship, this section describes the multiple compelling features that make PMGSY a favorable case for observing accountability.

Theoretical research suggests that voters are most likely to be electorally responsive to the provision of policies that are desirable, visible, easy to evaluate, and benchmarked (Mani and Mukand 2007). Evidence shows that rural road provision exemplifies these features (Harding 2015). Roads are highly desirable and important to Indian voters. Evidence from the largest pan-India voter survey (N=250,000) conducted in 2013 by the Association for Democratic Reforms indicates that voters rate roads as highly important with an average rating of 7.79 (scale of 0–10), closely behind employment (7.94) and drinking water (7.8). Indian media also documents citizens’ demand for roads and their positive response to PMGSY (see Hueiyen News Service 2012). Lack of access to roads is a recurring problem for which citizens are known to reach out to state politicians and seek help (see Hindustan Times 2018b). In areas that are poorly connected, citizens indicate the lack of roads as the reason for strong electoral punishment (see Telegraph India 2014). Some have issued cash incentives to politicians to visit their villages (see NDTV 2012). Citizens have literally taken matters into their own hands, carving roads to connect their villages to main roads (see Hindustan Times 2018a). Such events have inspired a blockbuster Bollywood movie called Manjhi—the Mountain Man, further increasing the salience of road provision. Politicians are aware of voters’ strong desire for roads. In the early 2000s most parties’ electoral campaigns promised roads, evidenced in the popular campaign slogan: “Bijli, sadak, pani,” or “Electricity, roads, water.”

Roads are visible and lead to large geographic changes in regions previously isolated. Because new roads are built where none existed before, it is much easier to benchmark the incumbent’s performance on road provision relative to challengers as well as relative to, for example, incremental improvements in school or road quality. Figure A2 presents a before and after image of villages that receive roads and illustrates the ease with which it is possible to benchmark road provision. Bhandari et al. (2023) find that information signals that can help voters to temporally benchmark incumbent’s performance are particularly effective. Specifically investigating citizens’ response to PMGSY, Sitapati (2014) further illustrates how voters are indeed able to temporally benchmark building of new roads under two different chief ministers (head of state government) in Bihar: “‘During Lalu’s times, I would fear my [car] suspension would be ruined by the mud roads, he says. I was worried I would be stopped and robbed. …’ Bittu’s greatest joy now is speed driving, ‘sometimes … even 70 [km per hour].’ This is because of Nitish, he adds. ‘He ended Lalu raj.’”

Evidence suggests that Indian voters attribute the responsibility for implementing the PMGSY to the state ruling party.4 India’s political context helps explain why: while central governments provide funding for large-scale programs, state governments implement them. Consequently, voters attribute responsibility for the provision of local goods and services to state ruling parties, reaching out to state legislators to benefit from development programs and schemes (Bohlken 2016). Evidence shows that quality of attribution is reasonably high in India (Goyal and Harding 2021), mainly due to the institutionalized and straightforward division of responsibilities between levels of government. Furthermore, ruling-party politicians influence service delivery because they have a high degree of leverage over bureaucrats (Gulzar and Pasquale 2017). Ruling parties’ stronger oversight over bureaucracy explains why voters attribute PMGSY to ruling-party politicians (Martin and Raffler 2021).

State governments were fully responsible for PMGSY service delivery. They could decide whether, when, and with what intensity they should implement the program. Consequently, not all states have been equally proactive in implementing PMGSY, with some exceeding their targets and some falling behind. Incumbent politicians and ruling parties promised rural roads in their campaigns; the media assigned their subsequent electoral success to the roads: “The excellent roads network was one of the principal vote-catching achievements of Nitish Kumar, who managed to beat anti-incumbency to return as chief minister in the state elections last year” (Economic Times 2016). Within each state, members of the state legislative assemblies (MLAs) jostled at district headquarters to influence bureaucrats to sanction roads in their constituencies. Bohlken (2021) documents political influence in the program; she finds that ruling-party MLAs have greater road length in their constituencies. Most crucially, MLAs influence program outcomes beyond the allocation of roads by also overseeing bureaucrats and contractors and ensuring that sanctioned roads are actually constructed.

To ensure politicians could claim credit for road construction, public foundation-laying ceremonies and inaugural town-hall-style meetings were organized with state and national politicians attending. PMGSY funded these ceremonies, and local media widely covered these events. Randomized audits helped ensure compliance with the rules that standardized how to conduct these ceremonies. Politicians used these events to claim credit for road provision (see fig. A1). Investigating policy attribution for PMGSY in Bihar, Sitapati (2014) finds that rural citizens have credited state governments for constructing PMGSY roads, as the following excerpt highlights.5 “Every single voter in West Champaran this reporter spoke to knew of the roads revolution, and credited the state government with it. Even Mantu Tiwari, a BJP supporter, grudgingly admits: ‘City roads were always fine, but yes. He (Nitish) has changed rural roads here.’ Traveling through the district on a burning afternoon, one sees girls in school dress running by freshly tarred roads, a sight unimaginable a decade ago.”

PMGSY also included transparency initiatives to ensure high-quality program information. From the program’s beginning, a distinct sign has marked each PMGSY road to differentiate it from other roads. The program specifies the placement and content of the standardized board, making it easy to spot PMGSY roads while traveling in rural India. Furthermore, each PMGSY road lists the central-level funding department and mentions the state-level government as being responsible for the program’s implementation. Internal quality data from PMGSY confirm that an overwhelming majority of roads (90%) have high-quality information boards.

Evidence indicates that citizens have a high recall for PMGSY roads and report daily usage. The Public Affairs Centre (2011) published a detailed citizen-led audit and monitoring report conducted in three different Indian states. This report finds that 94% citizen’s report a high level of perceived and felt benefits from PMGSY roads, and 84% claim to use these roads daily. The report also finds that a majority of citizens, ranging from 53% to 66%, recognize PMGSY roads and attribute them to state governments. Finally, causal evidence shows the positive impact PMGSY has had on variety of social outcomes, from access to health care to widening the marriage market and lowering crime (Adukia et al. 2020; Aggarwal 2018, 2021; Asher and Novosad 2020).

Data and Empirical Strategy

PMGSY roads-constituency data set

This section describes the process undertaken to build an all-India PMGSY roads-constituency data set. The first step was obtaining the road data. I obtained a consolidated PMGSY roads-level data set directly from India’s Ministry of Rural Development. This data set is not only almost twice as long in coverage years relative to all other studies investigating PMGSY (Asher and Novosad 2020), but it also contains the most comprehensive data on road quality. For each road, these data include all habitations (which are clusters of single-family dwellings; several habitations together form a village) and village names (and populations) a road connects; a road’s length, cost, funding date, construction start date, completion status, and date; and a quality rating for various aspects of each road. Although the program mainly built new roads, these data also contain roads that were upgraded and bridges that were built as part of the program. Data from fieldwork conducted during 2016–18 with the National Rural Roads Development Agency, the centralized bureaucratic department that oversees PMGSY, complement insights from the analysis of the PMGSY data. The fieldwork consisted of shadowing and conducting open-ended interviews with key program officials, attending quality assessment training, and visiting state-level public works departments in Jaipur and UP to interview state politicians, bureaucrats, and citizens.

The second step was adding census identifiers for each of the connected villages. To improve match quality, I merged the roads data set with the Census 2001 village data set using an exact and fuzzy name matching process for village names and population. The names were matched using a standard approach based on calculating Levenshtein distance and string similarities between canonicalized village names. On average, I could exactly match names and populations for over 76% of observations, while 16% were fuzzy matches. I keep Indian states where the total matching is at least 80%, the exact matching is at least 60%, and at least 2,000 villages are linked. This yields a sample of 15 large states, which contain close to 95% of India’s population. See table B1 for matching percentages for each state. Figure B1 plots the number of PMGSY projects. Note that the outcomes are measured at the constituency level. Each state constituency is geographically nested within administrative districts that are nested within states. National constituencies are typically as large as districts, but their boundaries do not overlap with those of districts. For each village in the 2001 and 2011 censuses, it is possible to add an identifier for constituencies and districts and therefore aggregate the results to the constituency level.

The third step involved adding state and national constituency identifiers to the roads-village data set and aggregating the data at the constituency level. A major border redistricting exercise was conducted in India in 2008, which makes electoral constituencies before and after this exercise incomparable. I refer to the period from 2000 until 2007 as “predelimitation” and from 2008 onward as “postdelimitation.” To add constituency identifiers for the predelimitation constituencies, I rely on data from Jensenius (2015). To add constituency identifiers for the postdelimitation constituencies, I rely on village and constituency maps from ML Infomaps. After adding the constituency identifiers, I estimated the percentage of villages that were connected by PMGSY within each constituency. I refer to this key predictor variable as “total connectivity” or “change in connectivity.” Figure 1 plots total connectivity within state and national constituencies for the pre- and postdelimitation periods. Roads are built to connect villages to one another and to the nearest economic centers. As such, as villages become more connected via roads, the constituency as a whole becomes more connected to other members as well (regardless of village population, to some extent). As a result, as a constituency becomes more connected, the likelihood that the incumbent is rewarded for facilitating this mobility increases.

Figure 1. 
Figure 1. 

Total connectivity across state (A, predelimitation; B, postdelimitation) and national (C, predelimitation; D, postdelimitation) constituencies. On average, PMGSY connected a substantive 8.3% of all villages (σ=7.8) in each state constituency during the predelimitation period and 9.2% of total villages (σ=9.8) in the postdelimitation period. Comparative data for the national constituency level are 6.9% (σ=5.6) and 8.3% (σ=6.4) of total villages connected in the pre- and postdelimitation period respectively.

The fourth and final step was merging this road-constituency data set with an electoral data set to add the dependent variable: the change in the state or national ruling party’s vote share. Because each constituency is observed exactly twice during pre- and postdelimitation, this variable is the difference in the vote share of the incumbent party or coalition in government between two consecutive state or national elections. There are no consolidated data on state-level incumbent governments. Therefore, I began by creating a list of consecutive state ruling parties for consecutive elections for both the pre- and postdelimitation periods. Figure B2 shows this list of ruling parties’ for each state. I used electoral data from Lokdhaba to estimate the change in each ruling party’s vote share. Appendix section B3 provides details about the data and variable construction. Given that the program was funded and launched by the national government, I also investigate whether voters reward national incumbents for road provision.6

Empirical strategy

Politicians anticipate electoral returns to PMGSY roads, which means that road provision is likely to be correlated with not only geographic or economic but also political considerations. To address these concerns and investigate the electoral returns to PMGSY connectivity, I estimate the following OLS regression:

where ΔYi is the change in incumbent vote share for constituency i, ΔXi is the change in PMGSY roads built or upgraded in constituency i, δj is a district fixed effect (all state constituencies are geographically nested within districts), and ε is a random error term. The coefficient of interest throughout the article is β, which captures the effect of road construction on the change in incumbent party’s vote share. Note that this is an FD estimation. However, because I consider only two time periods in any given panel, FD and fixed effects estimations yield identical results, as shown in Wooldridge (2013, 490). Following Wooldridge, I prefer FD estimations for two reasons. First, it is easier to interpret the point estimate, which is the change in incumbent vote share and aligns intuitively with the theory. It is also “easy to compute heteroskedasticity-robust statistics after FD estimation (because when T=2, FD estimation is just a cross-sectional regression” (490). Each constituency is observed exactly twice in each of the periods and all districts, and constituencies within a district/state undergo elections at the same point in time. In other words, there is no district-time or state-time variation.

This lean specification is similar to Aggarwal (2018, 2021) and Harding (2015) and robustly controls by design for (a) all time-invariant confounders at the constituency level such as geographic factors or preexisting infrastructure and (b) all time-variant and time-invariant confounders that vary across districts, for example, administrative capacity or district-level politics. However, because ruling-party politicians could partly influence roads allocation at the constituency level, roads allocation is still likely to be endogenous to within-constituency expectations. Existing research on distributive politics raises the expectation of positive electoral returns to road provision (Golden and Min 2013); studies documenting economic and social effects of PMGSY roads bolster this explanation (Adukia et al. 2020; Aggarwal 2018). The latter findings also allay concerns that the program was poorly implemented. Although less plausible, rent-seeking motives may override politicians’ electoral motivations to the extent that politicians target roads in constituencies with the most rent-seeking potential but no scope for electoral gains, increasing the odds of a null finding.

Endogenity concerns

To address concerns of political targeting, I use an IV approach that exploits the fact that PMGSY roads were provided partly along programmatic lines. An implementation rule targeted roads to villages with populations exceeding two discrete thresholds, 500 and 1,000, more generally, but 250 in hilly states, tribal districts, and districts affected by left-wing extremism. Table C1 lists the names of these specific states and districts. Village-level Census 2001 provided population data and was used to determine which villages are eligible. Figure 2A is a histogram of village population, and there is no visible sorting around the thresholds.

Figure 2. 
Figure 2. 

Population thresholds partly determine road provision: A, histogram of village population; B, probability of PMGSY connectivity by 2018. Plots include all villages that have a population between 1 and 3,000 as per Census 2001 (N=459,181). The mean village population is 855 (σ=693). The binscatter plot contains the default of 20 equally sized bins.

Each state could secure central funds to build roads in villages that qualified as per the implementation rule, which was one of the key factors in determining which villages to prioritize for connectivity. States also had discretion, and district-level administrations could propose villages that they determined required roads using local knowledge and economic importance. State-level offices compiled these district rural road plans and submitted the compiled proposal to the center to obtain the funding. States were responsible for allocating road contracts and implementing the program. Figure 2B shows that the implementation rules cause villages just above the population thresholds of 500 and 1,000 to be more likely to receive a road by 2018.

Figure 3A shows that villages just above the threshold of 500 are 8.9 percentage points more likely to receive a PMGSY project by 2018 than villages just below the threshold. This article is by no means the first to exploit these population thresholds as a source of exogeneity in road provision. Existing research in economics investigates the economic returns to roads at the village and district level, relying on the same underlying implementation rules for the purpose of identification. Aggarwal (2018, 2021) conducts a district-level analysis using an FD approach. Asher and Novosad (2020) conduct a village-level analysis by exploiting a fuzzy regression discontinuity (RD). These papers provide evidence that road provision is exogenous to village- and district-level characteristics. Crudely, the identification assumptions at the constituency level are stronger than those at the village level but less stringent than those at the district level, simply because constituencies are nested within districts.

Figure 3. 
Figure 3. 

Population thresholds and instrument. A, Probability of getting a PMGSY project by 2018 against village population centered at the cutoff of 500. Sample consists of villages within the optimal bandwidth of population thresholds (86). The optimal bandwidth in the full sample of villages is calculated using rdRobust. The point estimate for the discontinuity is 0.089 with standard errors of .015. B, Each unit is a state constituency in the predelimitation period with N=2,526. The mean of the eligibility instrument is 79.64 (σ=17.12). See table B2 for summary statistics for all levels and time periods.

Adherence to these implementation rules allows me to create an instrument that predicts road provision at the constituency level. This constituency-level instrument is the percentage of villages that are at or above 500 people in a given constituency.7 Figure 3B shows the distribution of the instrument for predelimitation state constituencies. Note that Census 2001 villages have been part of the same constituencies from 1977 through 2008, and the Indian Census is conducted by an independent centralized body that limits political influence. In other words, constituencies vary in their eligibility to receive PMGSY funds for reasons that predate the program and are exogenous to political influence. By estimating constituency-level FDs, I also control for time-invariant imbalances, which other studies have noted in PMGSY-eligible and noneligible villages (Aggarwal 2018, 2021; Asher and Novosad 2020). This FD IV approach is an improvement over Aggarwal (2018) but requires stronger identification assumptions than the fuzzy RD approach in Asher and Novosad (2020). I discuss these identification assumptions and threats to identification in the results section.

I estimate the following two-stage least squares regressions:

where ΔYi is the change in ruling-party vote share percentage in between two consecutive electoral periods; ΔRi is the percentage of villages connected by PMGSY in a given constituency between two consecutive electoral periods; Ei is the eligibility instrument, which is the percentage of villages over 500/250 population in a constituency as measured in Census 2001; δj is a district fixed effect, and ε is a random error term.

OLS Results

Table 1 reports the FD estimates obtained by regressing the change in vote share on the change in PMGSY connectivity during pre- and postdelimitation periods at the state and national levels. Increase in PMGSY connectivity has no effect on vote shares throughout the study period in both election levels. The point estimate is both substantively and statistically insignificant. The results are robust to alternative clustering techniques such as clustering at the district level (table B3). I also estimate these regressions in the subsample of constituencies where majority villages lack roads at baseline (table B4). Even in this subsample, where demands for roads is arguably higher, roads fail to boost electoral support for the incumbent. Results remain unchanged when analyzing a subsample of constituencies where challengers pose a viable threat in t0 measured as vote margin at or under 15% (table B5) or which are held by the ruling party in t0 (table B6). I also found that roads do not improve voter turnout in state elections, but there is a substantively positive effect in national elections (table B7). It is plausible that roads facilitated migrants to make long journeys and to return to vote in national elections that generally see lower turnout than state elections.

Table 1. 

OLS Results: Road Provision Uncorrelated with Change in Ruling Party(s) Vote Share

State ElectionsNational Elections
Δ connectivity.040.051−.097−.089
Mean Δ incumbent vote share−2.667−6.477−4.570−11.945
Fixed effectsDistrictDistrictStateState
Cluster SEConstituencyConstituencyConstituencyConstituency

Note. Dependent variable is the change in ruling-party vote share in state elections in cols. 1 and 2 and in national elections in cols. 3 and 4. Standard errors in parentheses.

*p < .05.

**p < .01.

***p < .001.

View Table Image

It is important to emphasize here that the OLS investigation favors the accountability hypothesis. OLS estimates rely on all of the variation in road provision that exists across the entire sample, including roads that politicians target electorally. Additionally, road provision increases citizen’s economic and social prosperity alongside multiple dimensions. Existing research on PMGSY shows that roads increase educational attainment for girls, health outcomes, marriage, employment, and access to loans. Yet, the effect of road provision on ruling-party vote share is substantively nil and statistically insignificant. These results run contrary to what existing models of voter accountability and economic voting would predict and are important in their own right. In the next section, I use IV regressions to show that the voters remain unresponsive to road provision that is not politically targeted but is instead determined by implementation rules. Together, this set of results strongly supports that voters did not reward (or punish) incumbents for road provision.

IV Results

Table 2 reports the IV estimates for the state and national elections in the predelimitation period.8 Change in connectivity shows a marginally negative relationship with the change in the ruling party’s vote share, but it is statistically insignificant in all estimations. Clustering the standard errors at a higher level of analysis (cols. 2 and 4) does not change the results.

Table 2. 

IV Results: Road Provision Does Not Increase Ruling Party(s) Vote Share

State ElectionsNational Elections
Δ connectivity−.380−.380−.264−.264
First stage: eligibility instrument.130***.130***.138***.138***
Fixed effectsDistrictDistrictStateState
Cluster SEConstituencyDistrictConstituencyState

Note. Dependent variable is the change in ruling-party vote share in state elections in cols. 1 and 2 and in national elections in cols. 3 and 4 in the predelimitation period. The eligibility instrument is defined as the percentage of villages in the constituency that are above the threshold of 500/250 population. Standard errors in parentheses.

*p < .05.

**p < .01.

***p < .001.

View Table Image

Threats to identification

Ordinarily, threats to identification are discussed in the context of there being a treatment effect, making it easier to think about the direction of the bias. In light of the null results versus the expectations of a positive finding, I weigh more heavily on identification threats that can bias the results downward.9 For the IV approach to serve as a plausible identification strategy, the eligibility of a constituency must have an influence on the outcome only through the PMGSY. While exclusion restrictions are an untestable assumption, several pieces of information suggest they are plausible. First, the centralized administration of the roads program determined these thresholds. A majority of the state and national politicians who were in office when the thresholds were established were no longer in office subsequently. Census data are also centrally collected by an independent and reputed government agency. Moreover, villages have largely remained within the same constituencies since 1977. Together, this means it is impossible for constituencies to select themselves into eligibility. Second, because the outcome is a change variable and not a level variable, time-invariant historical or preprogram imbalance, for example, geography or colonial administration, cannot explain the results. Third, the F-statistics for both instruments are always greater than 42 in the case of assembly constituencies and greater than 10 in the case of parliamentary constituencies. I also consider three other threats to identification that can bias the results downward.

Initial conditions and preexisting trends

Time-invariant, constituency-specific initial conditions cannot affect the outcome as they are controlled through the FD. However, they can still be correlated with underlying trends that can continue to affect electoral returns in eligible constituencies through mechanisms other than road provision. I investigate this threat by regressing changes in incumbent vote share and turnout in previous election years on the eligibility instrument and road connectivity. If voters have a history of being unresponsive or less likely to turn out to vote in eligible constituencies, it is plausible that service provision is not rewarded in such constituencies. I also include the change in BJP vote share for the state elections because the BJP formally launched the program and may have incentives to design the program eligibility rules in ways that are more likely to benefit its strongholds. Table C2 addresses concerns that the instrument is correlated with underlying pretrends in outcomes.

Figure C1 further shows that the instrument is uncorrelated with both previous and future employment and urbanization, which are likely to increase economic voting that favors the incumbent. Table C3 presents checks by adding change in employment and urbanization as controls in the two-stage least squares regression. The results are very similar, suggesting that change in economic conditions is uncorrelated with the instrument as well as the outcomes. Table C4 presents results in additional sample cuts where pretrends may bias results downward, such as constituencies that are already highly connected before the program start, constituencies with below average log employment growth overall, and constituencies in government and nongovernment sectors. I find that the results are substantively unchanged.

Thresholds predict other service provision

It is plausible that villages that became eligible for roads were prioritized or deprioritized using the same thresholds. I am not aware of any program that deprioritized villages using population thresholds. However, the Indian government launched an electricity and toilet provision program, where the eligibility criterion or incentives to implement these programs relied on population thresholds, in the later half of the 2000s. Consequently, villages in the predelimitation time period are no more likely to receive these other services. Unfortunately, the census is only conducted every 10 years, so I do not have outcome data that perfectly coincide with the predelimitation electoral period (before 2008) to explicitly test this. However, figure C2 replicates the connectivity binscatter in figure 2B, showing whether population thresholds in 2001 predict more households in a village are provided electricity or toilets in 2011. There is no clear visual jump in the probability of receiving these services as in the case of PMGSY connectivity. Very large villages (i.e., 75th percentile of population is 1,400) have a marginally higher probability of receiving electricity. The average number of households in rural India is 282 per village. A 5 percentage point increase reflects that approximately 14 more households are getting electricity or toilets in villages above 500 in population and is not substantively meaningful to drive the results in a positive direction. Moreover, the direction is opposite to how we expect the bias to operate. Note that the point estimate in the RD is also statistically insignificant.

Politicians underprovide other services in connected villages

Politicians know the program rules and which villages will qualify for the PMGSY. It is possible that politicians may systemically underprovide other crucial services in villages where PMGSY roads are provided and divert those resources to villages that do not qualify. Consequently, we observe that the effects of road provision cancel out due to the underprovision of other services at the constituency level, and this violates the exclusion restriction. Aggarwal (2021, 4) explores exactly this possibility and finds that PMGSY-beneficiary villages, over the 2001–11 period, “were no more likely than other villages to have received a school, a health centre, a railway station, or a bank branch.” I replicate this analysis with my data set and find that politicians do not underprovide other services in PMGSY villages, nor does PMGSY promote or undermine the delivery of other services.

Negative Spillover Effects

An advantage of an aggregated constituency-level analysis is that it accounts for positive spillover effects of the roads’ visibility. After all, citizens’ experience seeing new roads in their extended geographic area can lead those who do not get roads in their immediate community to punish the incumbent. In other words, negative spillover effects can occur. Aggregation may average out these microdynamics. To deal with this concern, I replicate the main OLS analysis at the level of the treatment: the village level. Investigating the change in polling station–level vote share within a constituency also enables holding politician characteristics constant.

I conduct this analysis in one of the most poorly connected and populous Indian states: Uttar Pradesh. UP has a population of more than 200 million and is geographically as large as the United Kingdom. Susewind (2014) provides geocoded data set for two postdelimitation state elections in 2012 and 2017; the only state where such data are available. Using data on approximately 109,217 polling stations situated within 389 (out of 403) assembly constituencies, I construct a PMGSY roads–Census 2001 village data set. I identify whether roads have been built within close proximity to a polling station. Lacking systematic information about which village votes in a given polling station, I overlay the geocoded location of the polling stations on village boundary maps based on Census 2001. I follow the literature to create spatial buffers around point polling stations and, if a road is provided in a village within a 1 kilometer radius, I assign that polling station as treated (Harding 2015).10 Table B7 shows that the results are insensitive to radius choice.

The analysis is highly localized: UP villages have a median population of 834. Between 2001 and 2017, approximately 95,000 kilometers of PMGSY projects were constructed, improving connectivity for 6% of all UP villages. Before delving in to the analysis, some specifics of the UP political context are relevant to the accountability relationship and worth noting here. Between 2012 and 2017, UP had a stable single-party majority (SPM) formed government—the Samajwadi Party (SP), a regional party—at the state level, making attribution easier for voters. Crucially, the SP won on a platform that spurned parochial identity-based divisions in favor of programmatic policies to improve development. It also faced a strong challenger, the BJP, which uprooted it from power in 2017.

Table 3 column 1 confirms that road provision had no influence on the SP’s vote share at the polling-station level, and the null result is precisely estimated. Column 2, 3, and 4 show that roads fail to boost SP’s vote share even in polling stations, which previously lacked roads in 2011, had no prior PMGSY provision, and lacked roads in 2001. Note that the point estimates are substantively marginal, and the large sample size makes it more likely that the null hypothesis is rejected by chance. Yet, the appendix shows that the results are mostly robust to clustering the standard errors at the constituency level, although they turn up significant when the buffer size is 2 kilometers.

Table 3. 

Roads Uncorrelated with Support for the Ruling Party at the Polling Station

PMGSY road.062−.121−.057−.073
SampleAllNo road in 2011No PMGSY road priorNo roads in 2001
Mean Δ SP vote share−8.417−9.167−8.485−9.135

Note. Unit of analysis is a polling station, and the dependent variable is the change in the SP vote share at that level. The independent variable PMGSY road is 1 if any of the villages within a 1 kilometer radius of the polling station gets a PMGSY project and 0 otherwise. Standard errors (in parentheses) are clustered at the polling station level.

*p < .05.

**p < .01.

***p < .001.

View Table Image

Why Accountability Fails: Mechanisms

The main results confirm that voters do not reward incumbents for providing roads. There can be several reasons why this accountability failure may occur. It is plausible that the null result masks heterogeneity in the quality of the new roads or corruption experienced by villagers (de Kadt and Lieberman 2020). In particular, some PMGSY roads have been marred by political influence in contracting, which may also lead to lower-quality road construction. Citizens may view high program costs as a signal of corrupt politicians who want to line their own pockets. PMGSY provides extremely rich road-level cost and quality data to test whether voters respond to the quality or cost of the new roads. I construct the following measures to tap into quality or corruption concerns: (a) log cost per kilometer of road length, (b) log maintenance cost per kilometer of road length, (c) log total cost per kilometer of road length, (d) average completion time in years, and (e) percentage of unsatisfactory road projects in relation to the total PMGSY projects undertaken in the unit. Table E1 reports the results. I find that there is no heterogeneity in voter response. Voters do not punish politicians for bad quality or costly roads at any political level, suggesting that corruption concerns are not the reason why voters punish incumbents (Klašnja 2015).

Roads are highly attributable within policy types (Harding 2015), and data from India and Ghana show that the quality of policy attribution across levels of government is high. However, it is plausible that the involvement of both state and national governments can complicate the attribution of roads in states with coalition governments or nonoverlapping state or regional parties (Powell and Whitten 1993). Building on the attribution literature, I identify subnational cases that have the potential to clarify politician’s responsibility to citizens. Constituencies with ruling-party politicians and states with SPM governments are expected to be more clearly responsible. SPM states aligned with central government also offer greater “vertical clarity” (Anderson 2006). These cases are Gujarat in the predelimitation period and Andhra Pradesh and Rajasthan in the postdelimitation periods. Additionally, both BJP and INC (Indian National Congress) are strongly organized in these cases, which further improves attribution (Jensenius and Suryanarayan 2022). Also note that these SPM states are highly competitive, and often the incumbent party loses elections in the next round. Table E2 reports estimates for ruling-party constituencies in the full sample, and in SPM states aligned with the central government. In both electoral periods, I observe either a weak or an inconsistent relationship between connectivity and incumbent vote share.

It is also plausible that voters only focus on road construction occurring close to elections and forget about roads that are provided early in the electoral cycle (Achen and Bartels 2016). Table E4 reports results showing that voters remain unresponsive to roads construction that happens close to election cycle in states where the first program cycle completed close to the election year. Table E5 shows that the rural voters in UP also do not respond to the building of new roads ahead of the election year.

Evidence suggests that information-media-rich and high-literacy environments in India can improve accountability (Besley and Burgess 2002) and lower incumbency disadvantage (Uppal 2009). To examine this possibility, I use village-level data on education and newspapers and media access aggregated to electoral levels to create constituency- and village-level measures of the extent of media richness. See tables E6 and E7. The estimates suggest that the building of new roads in media-rich environments does little to increase an incumbent’s vote share.

Finally, I investigate whether a high preference for challengers is the reason behind the weak accountability. To do so, I investigate whether a strong preference for challengers leads voters to reward the BJP, which can claim credit for PMGSY launch, only when it is the opposition party and presents a compelling alternative to the incumbent. See tables E10 and E11. The average change in BJP vote share is positive, which suggests that the constructed sample does capture the BJP as a compelling challenger. I find that road provision is positively and substantively correlated with change in BJP vote share in the majority of the specifications, particularly in the subsample of postdelimitation national constituencies, where it is more likely that BJP successfully engages in credit-claiming as challengers. Although the point estimate is statistically insignificant and results are preliminary, it is likely that road provision signals are more meaningful to voters when compelling challengers can claim credit for them. I also investigate whether constituencies where elections are highly competitive, and therefore campaigns more intense, are those where voters are more likely to learn about incumbent quality and respond to road provision. Tables E8 and E9 show that electoral competition does not moderate voter responsiveness. Table E12 replicates these results in UP.


Voters are electorally unresponsive to a multi-billion-dollar road construction program in India that connected close to half of India’s villages, most of which lacked paved roads and desperately needed all-year market access. These results are consistent across the whole of India, in different time periods, at different electoral levels, and in different units of analysis. India is also a typical case in the Global South, where existing studies raise mixed expectations about accountability. It is a low-income multiethnic democracy with a prevalence of ethnic and clientelistic voting and a strong anti-incumbency bias. Further, infrastructure development programs have become increasingly common in such settings, and, therefore, the findings are of direct relevance to research on the relationship between development and accountability. I expect the results will generalize to other more complex goods and services—for example, education and health care—as roads are a highly attributable public good (Harding 2015) and as such present a more likely case in which one may observe accountability. However, this remains an open empirical question.

This article moves the agenda toward addressing another key question: Why did voters not reward new road construction? The results in this article are compatible with the existence of substantial causal pathways between the building of new roads and electoral effects. However, the article rules out explanations that point to a breakdown in the informational chain as a key reason for the lack of performance-based voting. Findings suggest that attribution errors, concerns about corruption or quality, myopia, or the presence of a media-rich environment are unlikely to explain accountability failures.

The findings of this article have implications for the political agency model. It is still plausible that providing roads is rewarded or punished under conditions that question the very premise of political agency models, but these conditions are more likely to hinder accountability than enable it. For instance, these findings leave open the possibility that caste, religious, or partisan identities may shape how citizens evaluate new road construction or are able to coordinate with other voters to enforce accountability. Investigating each of these is beyond the scope of this article, but preliminary analysis suggests that it is likely that voters have a higher preference for challengers regardless of the incumbent’s quality (Weaver 2021). If this is true, providing voters with more, better, and unbiased information cannot enhance accountability; these findings echo other recent research that challenges the foundational premise of political agency models (Boas et al. 2021; Martin and Raffler 2021; Weaver 2021).

I thank Ben Ansell, Andy Eggers, Guy Grossman, Robin Harding, David Rueda, Francesca Jensenius, Bo Rothstein, Sam van Noort, Leonard Wantchekon, and Steven Wilkinson for their comments. Siddarth Venkatesh provided excellent research assistance in replicating the article. I thank bureaucrats at the National Rural Roads Development Agency for supporting the data collection in this article. I also thank the editors and three anonymous reviewers for their feedback. Previous versions of this manuscript were presented at the 2019 American Political Science Association and 2019 Midwest Political Science Association annual meetings, the 2018 14th Annual Conference on Economic Growth and Development at Indian Statistical Institute, Delhi, and the 2018 Oxford-LSE Graduate Political Economy seminar.


Tanushree Goyal () is an assistant professor of politics and international affairs, Department of Politics and Princeton School of Public and International Affairs, Princeton University, Princeton, NJ 08540.

Replication files are available in the JOP Dataverse ( The empirical analysis has been successfully replicated by the JOP replication analyst. An appendix with supplementary material is available at

1. De Kadt and Lieberman (2020) and Harding (2015) are notable exceptions to the rule when it comes to investigating accountability for public service delivery in Africa.

2. Studies that interpret voter behavior from any real-world electoral data, which are never observed at the voter level, require some level of ecological inference. Such inference is often necessary and is made by politicians and governments in electoral and policy analysis. Road access, in particular, is highly desirable but uneven and has been shown to improve voter prosperity. As a result, it is less problematic to anticipate that roads will have positive incumbency effects on the net, as discussed in de Kadt and Lieberman (2020) and Harding (2015).

3. Unlike in the case of relatively well-studied cash transfer programs and audits, corruption and endogeneity concerns have hindered progress on investigating voter responsiveness to public goods and services, limiting what we can learn from the few existing studies. De Kadt and Lieberman (2020) find that widespread corruption in South Africa has lead citizens to respond negatively to public goods provision, leaving it open whether voters will reward public services when both corruption concerns are less pronounced (or vary) and voters benefit from service provision. Harding (2015) is an exceptional study, finding that Ghanian voters reward incumbents for improvements in road quality, despite the prevalence of patronage or identity politics and even when politicians do not influence road maintenance. The findings bolster the expectation that voters are more likely to reward public goods provision in cases in which politicians do actually influence outcomes.

4. Depending on the policy of interest, it is plausible that voters may attribute responsibility to individual candidates. However, empirical evidence from the program suggests that the majority of voters credit the state’s ruling-party politicians for the PMGSY (e.g., see Public Affairs Centre 2011; Sitapati 2014). More broadly, evidence indicates that parties are the strongest determinant of vote choice in India’s state elections (Goyal 2019). For instance, in an exit poll election survey conducted by Lokniti in UP in 2017, only 10% of voters suggested that the individual candidate was the determining factor for their vote. Instead, 60% mentioned a party or chief minister, with the rest giving an inapplicable response. Other state election studies also find the party to be the most determinant.

5. This excerpt also highlights that voters can discern the difference in national (Bharatiya Janata Party, BJP) funding and program implementation by the state government, as confirmed in Goyal and Harding (2021).

6. Most voters do not attribute responsibility for PMGSY to individual legislators; rather, they attribute responsibility to state ruling-party politicians, as discussed in the context section. However, as a robustness check, I constructed the dependent variable as the change in the vote share of individual candidates and found that the results did not change.

7. This instrumental approach is not a fuzzy RD variant because road provision is a function of a constituency’s rule-based eligibility, but there is no discontinuity at the constituency level. Because reference groups cannot be designed around a threshold, the only possibility that remains is to construct an instrument that exceeds the threshold, which I do in this article. Undoubtedly, a fuzzy RD design presents the most compelling identification strategy. However, electoral data at the polling station level are unavailable during these time periods in state elections in India, except in UP. Unfortunately, states vary in the extent to which they follow implementation rules. For example, Rajasthan observed the rules more strictly, while UP ignored them. Consequently, I do not observe a discontinuity and cannot use a fuzzy RD approach in UP. Future studies can collect predelimitation polling station–level electoral data in states like Rajasthan. Despite several rounds of fieldwork, I was unsuccessful in collecting such data, partly because such data are not digitized or maintained for extended time periods.

8. I am unable to use the instrument for the postdelimitation period because program rules provide for fewer roads in this period, arguably due to the greater number of new roads constructed in the prior period.

9. For the same reason, I am unable to conduct tests that relax the exclusion restrictions as in Conley et al. (2012), as this test requires that there is a treatment effect.

10. A 1 kilometer radius buffer is reasonable given that Indian law requires that polling stations are within a walking distance of at most 2.5 kilometers.