Collective Choice and Project Delivery

While theories of distributive politics in developing countries have largely viewed expenditure decisions as the tactical choices of a unitary incumbent (Golden and Min Reference Golden and Min2013), in many contexts, expenditure choices are made not by a unitary incumbent but by a group of political actors who have different distributive preferences. The canonical scenario is a legislature that must collectively select a set of projects that each deliver benefits to only one constituency. No legislator has an incentive to support a project in any other constituency, but each needs the support of others in order to gain majority approval of their own project, and so legislators must assemble coalitions by agreeing to vote for each other’s projects. This is often referred to as “logrolling” over pork-barrel projects. Since not all projects can be implemented simultaneously and not all votes occur simultaneously—noncontemporaneous benefit flows and non-simultaneous exchange, in Weingast and Marshall’s (Reference Weingast and Marshall1988) terms—forming vote-trading coalitions requires intertemporal bargaining among legislators. But these intertemporal bargains—vote for my project this year, and I will vote for your project next year—are difficult to maintain, since legislators who have just received a project have an incentive to renege on their promise ex post. Knowing this, nonrecipient legislators have little incentive to sacrifice their own immediate self-interest (receiving a project in their constituency) in order to support another legislator’s project. These commitment failures make coalitions inherently unstable, as any structure of actor preferences that enables bargaining also enables voting cycles, making any existing coalition vulnerable to an alternative proposal (Riker and Brams Reference Riker and Brams1973).

Political bargaining over projects in developing country contexts such as Ghana’s local governments is typically not as formalized as logrolling in the US Congress, given the importance of informal politics and the relative weakness of legislatures, and may have de facto decision rules that are more complex or ambiguous than a simple majority vote. However, the project selection and delivery process in Ghana shares the key features of logrolling: spatially targeted projects that deliver concentrated benefits to one political actor but few or no benefits to other actors; the necessity of nonetheless gathering some degree of support from nonrecipient actors; and the time-inconsistency and thus instability inherent in the resulting intertemporal bargains. Senior Local Government Service officials also confirm the existence of these dynamics: while individual assembly members are primarily interested in projects in their own areas, formal and informal vote trading does occur as members seek broader support for their projects, and resources that had been allocated to one community are often reallocated later, leading to protests from the community.Footnote ¹⁴

While project implementation (as distinct from selection) is not a feature of existing formal models of collective choice, it is straightforward to see how the instability of collective expenditure priorities could affect development projects that are not instantaneously implemented. As Figure 2 illustrates, the technical and contractual process of constructing even small projects involves not just an initial phase of project selection and budgeting, but also an iterated process of construction (by the contractor) followed by payment (by the government) throughout the life of a project. A six-room school in Ghana, for example, has a median expected project duration of 5 months and may be divided into five payment tranches, the nonpayment of any of which can lead the contractor to stop work until payment is made. Since each disbursement on a project takes funds away from other potential projects, each payment tranche effectively requires a reaffirmation of collective expenditure priorities. The more unstable these collective priorities are, the more likely it is that a project will be interrupted with no guarantee when (or if) its completion will once again become a collective priority. The effect of this political instability is aggravated by technical and contractual characteristics of infrastructure projects: mid-project delays increase the cost of completing the project due to interest costs, physical decay in exposed works, and eventual relocation of the contractor’s staff and plant.Footnote ¹⁵ Since the contract sums of small development projects are typically not indexed to inflation or exchange rates, these cost increases can render further work on contracts unprofitable for contractors after even relatively short delays.

FIGURE 2.

Stylized Timeline of a Small Infrastructure Project

FIGURE 3.

Project Type by Fund Source

While novel to the political economy literature, the idea that unstable political support coalitions can affect project implementation features prominently in the qualitative policy implementation literature within public administration. For example, Pressman and Wildavsky (Reference Pressman and Wildavsky1984 [1973], 92) note in their classic study of urban development projects in Oakland that “Agreements [among project stakeholders] were reached, eroded, and remade. The frequent calls for coordination. . .reflected the inability of the machinery for implementation to move fast enough to capture the agreements while they lasted.” Similarly, political actors in Ghana express concern that the government’s ability to efficiently provide public goods is undermined by its inability to simultaneously satisfy all the distributive demands on it. This is perhaps most clearly illustrated in a speech by the President of Ghana, John Mahama:

The President’s invocation of a familial metaphor to explain the necessity of sequencing project delivery, and his pleading with voters for patience “in order that we do not disrupt the budget and government expenditure,” concisely summarizes both the necessity and difficulty for politicians of intertemporal bargaining over the delivery of public goods, as well as the negative efficiency consequences of failing to maintain these agreements. District-level actors also express similar concerns about how distributive pressures can undermine project execution. At a public hearing about a project audit, the bureaucratic head of Sekyere South district explained that the district bureaucracy “implements recommendations from the General Assembly. They have reduced projects but the citizens keep demanding projects. . .They [the General Assembly] try to prioritise but it is not easy to do so” (Ghana Audit Service 2016b, 68). A senior Audit Service official involved in project audits explained the consequences of having to pay for the constant stream of new projects: “It’s not shown that it’s being taken from the budget, but what it means is that the other projects won’t be completed. And unfortunately these districts, a lot of the projects that they start they don’t complete.”Footnote ¹⁶

Observable Implications

These collective choice dynamics cannot be examined directly in my empirical case, as much of the bargaining is informal and there is no data on proceedings and voting in district assemblies in Ghana (as is the case with local-level governance in many developing countries). I instead identify the impact of collective choice failures indirectly, by making theoretical predictions about characteristics of districts and fund sources that are likely to make sustaining intertemporal bargains over project distribution more (less) challenging, and thus lower (raise) completion rates.

Legislative institutions such as committees that enable agenda-setting and facilitate enforcement through rewards or punishments are key to resolving problems of the durability and enforceability of legislative bargains (Weingast and Marshall Reference Weingast and Marshall1988; Baron and Ferejohn Reference Baron and Ferejohn1989). In the Ghanaian context, the executive is more likely to be able to stabilize local project support coalitions in districts that are politically aligned with the ruling party, due to two asymmetries in the structure of district governance in Ghana: the DCE is always appointed by the President, regardless of district vote shares; and in every district, 30% of assembly members are appointed by the President to at-large seats.

DCEs can use their powers of ex ante agenda-setting—through their control of the DPCU, their chairing of the powerful Executive Committee, and their influence on committee memberships—and ex post implementation of assembly decisions to try to ensure that started projects get finished. However, their incentives to do so are much stronger when these projects benefit the ruling party—projects which incumbent-leaning assemblies are more likely to have started in the first place. Since DCEs control many nonproject aspects of expenditure, they can also use other side-payments to induce cooperation from assembly members. Not only are they likely to be more willing to do so with respect to members of their own party, but they may also be better able to commit to doing so due to the availability of intraparty sanctioning mechanisms. The Presiding Member of the assembly can also help stabilize resource allocations through their committee appointment powers, but this power is likely to support the DCE’s efforts when they are co-partisans and work at cross-purposes when they are not.

There is ample qualitative support for the idea that DCEs selectively use their formal and informal powers to coordinate with co-partisans on project delivery. Local Government Service officers confirm that DCEs and opposition assembly members often have different preferences on project locations, stating that “DCEs in opposition strongholds will have a fight” to agree on project locations with the assembly.Footnote ¹⁷ Similarly, Nathan (Reference Nathan2016, 8) quotes one NDC assembly member as explaining, “[The DCE is] a government appointee and I’m also from the same government. . .He listens to the NDC assembly members more than the others. It’s the norm,” while in contrast, an NPP assembly member lamented, “But then your project, you have to spend years chasing it before you get the money. . .I don’t bother going to the [DCE’s] office. . .because the [DCE] has his own agenda, his own political agenda. . .The other assemblymen will have finished their projects, and they’ll be saying no money for you.”

TABLE 1.

Observable Implications of Theories of Noncompletion

Likewise, whereas the electoral incentives and thus project distribution preferences of elected assembly members are divided between bringing projects to their own electoral area and supporting projects in other areas that would be electorally beneficial to their party, the incentives of the appointed assembly members are to support whatever projects will be electorally beneficial for the incumbent. This is consistent with qualitative accounts of appointed members’ partisan affiliations, with (for example) a former president of the National Association of Local Government Authorities of Ghana stating, ‘About 95 percent of the appointees to the assembly by the President belong to the party of the President” (Citi FM 2016). Appointed assembly members can thus act as a stabilizing force for project support coalitions of elected members, but have stronger incentives to do so when these coalitions are comprised of members of their own party.

Project completion should therefore be positively correlated with the district’s partisan alignment with the ruling party, which I proxy by the ruling party’s vote share in the presidential election preceding the study period.Footnote ¹⁸ However, a simple test of this prediction using cross-district variation in partisan vote shares and project completion could be biased due to the potential endogeneity of voting patterns and public good provision or to unobservable confounds more generally.

Instead, I exploit within-district variation in the completion of projects across different fund sources. As discussed above, the donor-funded transfer program’s rule that districts must budget to finish ongoing projects before starting new ones makes donor-funded projects less vulnerable than government- funded projects to collective choice failures by strengthening nonrecipient assembly members’ incentives to complete ongoing projects located outside their electoral area. The within-district difference between donor- and government-funded project completion rates should thus be an indicator of the impact of collective choice failures on project completion.Footnote ¹⁹ This funding rule is likely to have the strongest effect in districts that are politically aligned with the opposition party, where these collective choice failures are most acute.

The key observable implication of the collective choice theory is therefore that the completion rate of government-funded projects should increase in districts with higher ruling party vote share relative to the completion rate of donor-funded projects in the same district. Since this prediction relies on within- rather than across-district variation in completion rates, it is possible to control for the observable and unobservable characteristics of districts, as well as for the politicians and bureaucrats selecting and implementing the projects, thus eliminating concerns about endogeneity and unobservable district characteristics.

A further implication of the idea that the ruling party’s vote share in a district is a proxy for district-level collective choice failures (rather than a characteristic on which the incumbent tactically targets completion) is that vote share should have no effect on the completion of centrally delivered projects in the district. The completion rate of centralized projects relative to the completion rate of donor-funded projects should therefore be unrelated to district ruling party vote share.

Alternative Explanations: Corruption and Clientelism

There are two other plausible explanations for project noncompletion in which noncompletion is a deliberate outcome rather than a collective choice failure. First, noncompletion could be caused by a specific form of corruption: projects could be left incomplete if contracts were given to favored contractors who are paid for the work but placed under little pressure to complete it. This would enable the contractors to keep as profit, or pay as kickbacks, the cost savings from not completing the project.Footnote ²⁰

While this type of corruption on the “quantity” margin—as opposed to the quality and price margins that the existing literature has focused on (Olken Reference Olken2007; World Bank 2011)—has not previously been documented, it does have a clear empirical implication: financial expenditure on noncompleted projects should be greater than the physical progress of construction on these projects. In effect, this test is analogous to Olken’s (Reference Olken2007) well-known indicator of “missing expenditures,” albeit where the “missing” expenditure is on the quantity rather than the quality margin. In contrast, an observable implication of the collective choice theory is that the “missing expenditures” measure described above should be negative, reflecting underpayment to contractors for work done, as projects are interrupted mid-construction by nonpayment due to shifting collective expenditure priorities.

Second, noncompletion could be the outcome of a clientelist strategy by the ruling party to use project completion decisions as a way to reward or punish voters. However, existing theories of clientelism as tactical targeting of voters overwhelmingly focus on the allocation of projects, not their completion, so on their own, these theories do not explain why a government would choose to allocate its scarce resources to begin a project in a particular community, but then not finish it. This observation is not to challenge the prevalence of clientelism in project allocations or voting patterns in Ghana or elsewhere, on which there is a large and well-established literature (cf. Golden and Min Reference Golden and Min2013), but rather to distinguish theories of the distribution of projects (which is not my focus) from theories of the implementation of these projects conditional on them having been started (which is).

Nevertheless, one way in which clientelist dynamics could lead to project non-completion is if politicians deliberately leave projects incomplete before elections due to voters’ inability to credibly commit to reward projects with votes. This alternative version of the clientelism theory builds on the logic underlying theories of clientelism as a two-way credibility problem (Robinson and Torvik Reference Robinson and Torvik2005; Robinson and Verdier Reference Robinson and Verdier2013). While project completion is not the focus of these models, the logic of two-way credibility problems—that politicians sometimes find it rational to make apparently inefficient public expenditure choices in order to “tie the continuation utility of a voter to their political success” (Robinson and Verdier Reference Robinson and Verdier2013, 261)—could potentially be applicable to project completion. If credit-claiming dynamics make the incumbent more likely than a challenger to complete a project started by the incumbent, starting but not finishing a project would increase nearby voters’ incentives to vote for the incumbent, and so immediately before an election, incomplete projects may actually be more desirable from the incumbent’s point of view than completed projects.

This clientelism-based theory of noncompletion has several observable implications. First, project noncompletion should vary significantly across the electoral cycle. Since Ghana held a presidential election in December 2012, the clientelism theory would imply a lower completion rate in the election year (2012) than in pre-election (2011) or postelection (2013) years.Footnote ²¹ A second implication is that the degree of physical progress on started projects is likely to be low, since the project start itself is what creates the desired voting incentives (from the incumbent’s point of view), and further expenditure on the project does not enhance these incentives. Third, and in contrast to the predictions of the collective choice theory, if the ruling party views project noncompletion as an effective political strategy then the completion rate of centralized projects should mirror that of government-funded projects—i.e, should be positively correlated with ruling party vote share relative to donor-funded projects.Footnote ²²

Finally, there could also be other causes of project noncompletion that are outside the scope of this article, either because they are not directly related to political economy dynamics or because they are not relevant in my empirical context. Examples of the former could include the competence of bureaucrats and contractors, geographical or economy characteristics of communities, district budget sizes, or simple bureaucratic turnover, which I am able to fully control for empirically, and thus should not bias the analysis. Examples of factors that are not relevant in my context could include elections that lead to the incumbent’s removal or severe economic shocks that lead to unanticipated fiscal contractions, although it is easy to see how these could also cause project noncompletion in other contexts. I discuss the scope conditions of this article’s theoretical argument further in the conclusion.

Data

To measure the completion of infrastructure projects, I collected, digitized, and coded all available district Annual Progress Reports (APRs) for the years 2011–13. These reports are written annually by each district’s bureaucracy and submitted to the central government, and include a table listing basic information about projects that were ongoing or active during the calendar year (Appendix A). Altogether, it was possible to locate 479 out of a potential 602 district-year APRs (79.6%), of which 407 (67.6%) had sufficient information to be entered into the database.Footnote ²³ After cleaning and removing noninfrastructure projects, this yielded a database of 14,246 project-year observations, which is the core dataset used in this study’s analysis.

Tracking the progress of projects across years is challenging, since many projects do not have commencement dates reported, few districts make use of unique project identification numbers, data is censored before 2011 and after 2013, and many incomplete projects attrite out of the dataset. I use three alternative methods of linking project observations across years and take numerous steps to verify that the key results are not being driven by reporting bias, attrition, or double counting; these are reported in Appendices D, E, and F. Most importantly for my empirical strategy, Appendix D shows that attrition is uncorrelated with project fund source and thus should not bias my main estimates.

All reports were manually double entered into the database and variables were coded through text matching from these strings, with manual disambiguation for key variables (see Appendix C). Projects were coded into 17 different basic “types,” such as schools, clinics, and staff housing. The category “schools” was broken down into five subcategories according to the number of classrooms, leaving a total of 22 project type categories. Project completion was coded for each observation as a binary variable by combining information from three raw variables. For instance, project completion was coded as 1 for values such as “COMPLETE,” “100%,” or “INSTALLED AND IN USE,” and 0 for values such as “ONGOING,” “90%,” or “LINTEL LEVEL.”Footnote ²⁴ I validated the accuracy of districts’ completion reporting in several ways, including site visits, comparison against smaller datasets on project completion compiled through independent evaluation reports, interviews, and alternative government data sources, and found no evidence of systematic bias or misreporting of completion status in the reports (Appendix E).

In addition, for 140 districts, I was also able to digitize District Medium-Term Development Plans. These were 3-year plans written in late 2010 that included tables of planned development projects through 2013, and for the subset of districts with no missing APRs, I linked these planned projects (using the same procedure as linking across years of the APRs) to records of actual projects undertaken over the plan period from the APR dataset.

Extent, Dynamics, and Costs of Noncompletion

Less than half of projects (45.8%) were finished within their first year of implementation, and even after three years, 35.5% remained unfinished.Footnote ²⁵ These low completion rates were not due to long completion schedules: the median project had a scheduled duration of five months and 88.8% were scheduled to be complete within twelve months. The completion hazard rate decreases after the first year of a project: conditional on not being complete after one year, over half of projects see zero or near-zero physical progress in the subsequent year.Footnote ²⁶ A similar bifurcation of project outcomes is evident when examining delays: the median completed project was delayed by just one month, 81.0% were finished in less than twelve months, and just 10.7% were delayed by a year or more. The median incomplete project, however, is twelve months past its expected completion date—a 200% delay—and there is a long tail of over a quarter of incomplete projects that are more than two years past their planned completion date. These dynamics are consistent across project types, as Figure 4 shows for the three most common types.

FIGURE 4.

Project Completion

While the availability of only three years of data makes it impossible to directly observe the eventual completion rate of projects left incomplete after three years, 99.5% of projects were scheduled to be completed within three years, and examining the completion dates of completed projects reveals that just 2.9% of completed projects took three years or more to complete. Assuming that the past distribution of project completion times reflects the time-to-completion distribution of projects started in a given year, extrapolating this rate out from the observed three-year completion rate implies that 33.5% of all projects are never completed.Footnote ²⁷ Of course, this estimate is based on the available three-year window of data that may or may not be representative within Ghana, and thus project completion rates could be lower or higher in other time periods (or other countries).

A key observable implication of the clientelism theory of noncompletion is that project completion rates should vary across the electoral cycle, and be lower in the election year, 2012. However, Figure 5 shows that there is no statistically significant variation in completion rates across years for government- or donor-funded projects. Completion rates are actually slightly higher in the election year for centralized projects for which the central government is responsible for implementation, suggesting that—if anything—the incumbent views completing projects just prior to elections is more electorally beneficial than leaving them unfinished. While electoral cycles could affect project completion in other contexts, particularly if the incumbent loses, in this case, the phenomenon of unfinished projects is remarkably consistent across the electoral cycle.

FIGURE 5.

Noncompletion Not Driven by Electoral Cycle

Most unfinished projects have had a significant amount of work done on them: projects that remain unfinished after their third year average 64.9% physical completion and have seen 55.5% of the contract sum disbursed to the contractor. The heavy investment into unfinished projects contradicts a key observable implication of the clientelism explanation for noncompletion, and also illustrates the fiscal and social costs of noncompletion. If 33.5% of projects are never finished and expenditure on these averages 55.5% of the contract value, a back-of-the-envelope calculation suggests that 18.6% of total local government capital expenditure is spent on projects that will never be finished. Scaling this by the total capital expenditure of Ghana’s local governments in 2013 of US$135 million implies annual spending on never-to-be-completed projects of approximately US$25.1 million. If this wasted expenditure were to be spent entirely on new three-room school buildings, it would be enough to fully construct 667 additional schools per year, so the social opportunity cost of noncompletion is also substantial.Footnote ²⁸

To examine the observable implications of the corruption and collective choice theories of noncompletion related to over- and underpayment of contractors, I construct a two-sided analog to Olken’s (Reference Olken2007) missing expenditures measure by subtracting the percentage of physical construction that has been completed on a project from the percentage of the contract’s value that has been paid by the district to the contractor. Positive values would be consistent with corruption in procurement—either through kickbacks or by selecting politically connected contractors who are paid but not put under pressure to finish projects—as a cause of noncompletion, while negative values would be consistent with nonpayment of contractors due to districts’ unstable collective expenditure choices. Of course, positive values may also be due to factors other than corruption, such as project cost overruns, so this measure likely overstates the extent of corruption-related overpayment. While the relationship between these values need not be strictly linear, depending on the details of contracts and construction, for these small projects, there are unlikely to be large discontinuities, and for completed projects, at least they should be equal. Appendix E discusses steps taken to validate these completion measures and shows that underpayment is not explained by normal time lags in processing payment.

Empirically, Figure 6 shows that underpayment of contractors is far more common than overpayment. For incomplete projects, underpayment by 10% or more is over three times more frequent than overpayment by 10% or more, and for completed projects underpayment is over sixteen times more common than overpayment. This is the opposite of what would be observed if corruption on projects were a significant driver of noncompletion, but is consistent with project interruptions driven by unstable collective priorities: projects start, the contractor does part of the work, but at some point during construction, the government’s collective priorities shift and the government thus fails to make a payment for work done, so construction stops.Footnote ²⁹

FIGURE 6.

Missing Expenditures and Underpayment for Completed Work

The prevalence of underpayment is also consistent with qualitative accounts of frequent complaints by contractors in Ghana about delayed and nonpayment by government (Abotsi Reference Abotsi2013) and with the conclusions of central government auditors (Ghana Audit Service 2016a; see Appendix E). The link between districts’ failure to pay contractors for work done and districts’ shifting collective expenditure priorities has also been noted by auditors: a 2013 report of districts’ spending of statutory funds criticized “the failure of the Assemblies to complete and pay for projects before embarking on new ones” and “recommended that management of the Assemblies should endeavour to discharge all obligations on time and also ensure completion and payment of ongoing projects before commencing new projects” (Ghana Audit Service 2014, 19).Footnote ³⁰

Finally, comparing planned projects to actual projects reveals a high degree of instability in district expenditure priorities. For the 39 districts with digitized medium-term plans and all three years of APR data, on average, only 5.7% of projects that districts planned for the period were actually started, and only 3.8% of projects reported in the APR database could be identified in the plan for the period.Footnote ³¹ These patterns are consistent with the premises of the collective choice theory that there is a large number of potential projects for which demand exists, and a great deal of temporal instability in which projects are recognized as collective priorities at any given moment.

Empirical Strategy

To examine the remaining observable implications of the collective choice and clientelism theories, I estimate a linear probability model (LPM) of the form:

(1) $$\begin{eqnarray} y_{i,j,k,t} &=& {\bm{FS}_{i,j,k,t}}{\bm{\beta} } + {\bm{FS}_{i,j,k,t}}*V_{k}*{\bm{\tau}} + {\bm{P}_{i,j,k,t}}{\bm{\gamma} }\nonumber\\ && +\, \lambda _{j} + \theta _{k,t} + \varepsilon _{i,j,k,t} \end{eqnarray}$$

where y _{i, j, k, t} is a binary indicator of completion of project i of type j in district k and year t,Footnote ³² FS _{i, j, k, t} is a vector of fund source indicator variables (government-funded and centralized, with donor-funded the omitted category), V _k is the vote share in the district of the ruling party during the study period (the NDC) from the first round of the 2008 presidential electionsFootnote ³³ , P _{i, j, k, t} is a vector of project characteristics including construction type (i.e., new construction vs. maintenance or repair) and the number of years since project start, λ _j is a vector of project type fixed effects, θ _{k, t} is a vector of district-year fixed effects, ε _{i, j, k, t} is an error term, and β , γ , and τ are vectors of parameters to be estimated.

The parameters τ are of primary analytical interest. The parameter τ ^Gov't represents the marginal effect of partisan vote share on government-funded project completion relative to the marginal effect on donor-funded project completion. A positive τ ^Gov't would be consistent with the key observable implication of local-level collective choice failure in project completion. The identifying assumption is that any correlation between the within-district donor-government completion rate gap and the district’s partisan alignment is driven only by the collective choice mechanism discussed above, so other factors could cause differences between the overall completion rates of donor- and government-funded projects without biasing my analysis so long as they are uncorrelated with district vote share V _k . Selection of projects into fund sources on unobservable project characteristics is the major concern with this assumption’s validity, but I show below that such selection effects appear to generate, if anything, a downward bias on τ ^Gov't and thus understate the true effect.

In turn, the parameter τ ^Central shows the marginal effect of district partisan vote share on centralized project completion relative to its effect on donor-funded project completion. This interaction term functions similarly to a placebo test, in that if the effect τ ^Gov't is driven by district-level collective choice failures, then τ ^Central should be zero since these district-level collective choice processes should not impact project completion decisions by the central government. On the other hand, if τ ^Central is also positive, then τ ^Gov't could be reflecting a deliberate clientelistic strategy by the government, which it is somehow constrained in employing with donor-funded projects.

Interaction of Fund Source and Partisan Alignment

I begin by estimating a simplified version of Equation (1) (Table 2, Column 1) on the sample of government- and donor-funded projects, with no fund source-vote share interaction term and with district-year random effects rather than fixed effects to enable the inclusion of the time-invariant variable V _k (NDC vote share). As expected, the coefficient on the Government-funded dummy is negative—government-funded projects have a conditional annual completion probability 10.9 percentage points lower than that of donor-funded projectsFootnote ³⁴ —and the coefficient on NDC vote share is positive, although not statistically significant. This direct test of the effect of NDC vote share is potentially misleading for two reasons, however: first, it assumes that this effect will be the same for both government- and donor-funded projects; and second, it does not fully control for district characteristics that could be correlated with both project completion and NDC vote share.

TABLE 2.

Interaction of Project Fund Source and Partisan Vote Share

Note: Dependent variable is project completion. All specifications include: fixed effects for 22 project types, with schools grouped by number of classrooms; construction type (construction or maintenance); and indicators for number of years since project start. Year fixed effects included in Columns 1-2 and 6. NDC vote share 2008 is the ruling party’s vote share in the 2008 presidential elections in the district (or its antecedent district, for districts that split in 2012). Huber-White robust standard errors clustered by district-year in Columns 1-4, by contractor in Column 5, by community in Column 6, and by community-year in Column 7.

To address the first issue, Column 2 adds the Gov’t-funded * NDC vote share interaction term. Consistent with the collective choice theory, the effect of NDC vote share is entirely driven by government-funded projects: $\hat{\tau }^{Gov't}$ is positive and statistically significant (p = 0.031), while the main effect disappears entirely.

Column 3 presents the full model in Equation (1), using district-year fixed effects to fully control for all observable and unobservable district characteristics, both time-variant and -invariant.Footnote ³⁵ The coefficient $\hat{\tau }^{Gov't}$ remains positive and statistically significant (p=0.035). The estimated effect is also economically important: a one standard-deviation increase in the ruling party’s 2008 vote share in the district (16.9 percentage points, from a mean of 49.7) is associated with a 4.36 percentage point decrease in the donor-government completion rate gap.Footnote ³⁶ This is equivalent to closing 39.8% of the total completion gap between donor- and government-funded projects. In absolute terms, a new six-room school in a district at the 5^th percentile of NDC vote share (corresponding to a strongly opposition-supporting district where the NDC received only 21.1% of the vote) would have an estimated annual completion probability of 49.7% if it were donor-funded (CI: 38.9-60.4) but just 32.4% if government-funded (CI: 22.4–42.4), a gap of 17.3 percentage points. In a district at the 95^th percentile of NDC vote share (88.9%), however, this gap disappears: the government-funded school would have a completion probability of 47.6% (CI: 36.1–59.1), almost identical to that of the donor-funded school (47.8%, CI: 34.6–60.9).Footnote ³⁷ This is strong evidence in support of the collective choice theory of noncompletion: in districts where local political actors are almost uniformly aligned with the incumbent, the executive is better able to use its agenda-setting, implementation, and intraparty enforcement mechanisms to uphold intertemporal bargains on project sequencing and thus avoid leaving projects incomplete.

To further discriminate between the observable implications of the collective choice and clientelism theories, Column (4) adds centralized projects into the estimation sample and introduces the centralized dummy and interaction term. While $\hat{\tau }^{Central}$ is not significantly different than zero (indeed, is slightly negative), $\hat{\tau }^{Gov't}$ remains positive and significant, and the difference between the two coefficients is statistically significant (p = 0.017). This is inconsistent with the clientelism theory that government-funded project completion patterns reflect a deliberate policy of the incumbent either to leave projects unfinished in order to give voters an incentive to return them to office or to punish opposition voters. If the government thought this to be an effective political strategy, we would have expected that NDC vote share should have a similar effect on centralized projects as it does on government-funded projects executed by districts. Instead, the markedly different effects are consistent with the collective choice theory that the incumbent is constrained in the completion of government-funded projects by local-level collective choice problems that are more difficult to resolve in opposition-leaning districts. The district-year fixed effects in this specification also exclude the possibility that the observed effects are driven by changes in districts’ political leanings over the study period, which is further evidence that these results are not driven by clientelist dynamics of using project completion tactically to reward or punish voters. Figure 7 presents the marginal effects of NDC vote share on the three fund sources graphically, with donor-funded projects as the baseline category represented by the dashed horizontal line at zero.

FIGURE 7.

Marginal Effect of Partisan Alignment by Fund Source

A further concern is that the relevant clientelist dynamics might operate not at the level of whole districts, but rather across communities within districts. For example, if within each district the government systematically allocated donor-funded projects to its core supporter communities and government-funded or centralized projects to opposition communities, and if donor-projects were more likely to be completed and projects in core supporter communities were also more likely to be completed, then the estimates of $\hat{\tau }^{Gov't}$ in Columns 14 could be biased upward. To address this, Column 5 re-estimates Equation (1) with fixed effects for the 2,934 communities in which projects were executed, to control for all observable and unobservable characteristics of these communities, including partisan affiliations as well as other more general community-level determinants of project completion. Controlling for these community-level factors actually slightly increases the point estimate of $\hat{\tau }^{Gov't}$ , although the statistical significance drops slightly (p = 0.077) due to an increase in standard errors. Column (6) pushes this even further by controlling for community-year fixed effects, thus capturing any shifts over time in partisan affiliations that might affect project completion, leading to $\hat{\tau }^{Gov't}$ increasing even further and becoming more statistically significant (p = 0.034). The continued significance of district-level partisan affiliation in explaining government-funded (but not donor-funded or centralized) project completion—even after controlling for time-variant community-specific determinants of project completion—is strong evidence that these results are driven not by clientelist strategies of using project completion as a reward or punishment, but rather by district characteristics that affect collective choice outcomes.

Finally, Column (7) adds contractor fixed effects, since $\hat{\tau }^{Gov't}$ could be biased upward by contractor identity if the donor-government completion gap was driven by contractor quality and less able contractors were more likely to be selected in opposition-leaning districts. This specification takes advantage of the fact that some of the 2,592 contractors who implemented these projects did so in multiple districts and/or for multiple fund sources. The estimated $\hat{\tau }^{Gov't}$ actually increases slightly and remains statistically significant (p = 0.050). This provides further evidence that the observed patterns of project noncompletion are not being driven by theories of corruption and clientelism that focus on rent-seeking behavior in procurement by politicians, bureaucrats, and politically connected contractors.

These results are robust to a wide range of alternative specifications and additional controls; Appendix F presents these specifications.

Addressing Endogenous Selection into Fund Sources

Even though the district-year, project type, community, and contractor fixed effects control for a wide range of observable and unobservable variables that could bias the analysis, one remaining concern is that communities may be sufficiently heterogeneous that projects could even be targeted (i.e., be locally excludable) within communities, for instance, based on very local ethnic or partisan geography (Ichino and Nathan Reference Ichino and Nathan2013). If projects are locally excludable and thus can be targeted based on within-community unobservable variation, this could potentially bias $\hat{\tau }^{Gov't}$ upward if the extent of within-community selection of projects into fund sources was also positively correlated with the ruling party’s vote share in the district.

Although I lack detailed geographic data with which to measure this within-community heterogeneity directly, I am able to gauge whether this bias could plausibly be driving $\hat{\tau }^{Gov't}$ using the coefficient stability approach developed by Altonji, Elder, and Taber (Reference Altonji, Elder and Taber2005) with corrections from Oster (Reference Oster2016) and Gonzalez and Miguel (Reference Gonzalez and Miguel2015). Oster shows that if selection on unobservables (such as within-community variation) is as strong as selection on observables (across-community variation), a bias-adjusted estimate of τ is:

(2) $$\begin{equation} \tau ^{*} = \tilde{\tau } - (\dot{\tau }-\tilde{\tau })\frac{R_{max} - \tilde{R}}{\tilde{R}-\dot{R}}, \end{equation}$$

where $\tilde{\tau }$ is the estimated coefficient from a regression including community fixed effects (Column 2) and $\tilde{R}$ is the R ² from this regression, $\dot{\tau }$ is the estimated coefficient from the regression excluding community fixed effects (Column 1) and $\dot{R}$ is the R ² from this regression, and R _max is the percentage of all variation in the dependent variable theoretically explainable by all possible observables and unobservables. Even under the most conservative case of R _max = 1 where within-community unobservables account for all remaining unexplained variation and there is no measurement error, calculating τ* ^Gov't using Equation (2) and the estimates from Table 3 still generates a bias-adjusted coefficient of 0.253. More intuitively, these estimates imply that the process of endogenous selection into fund sources on unobservable within-community variation would have to be 58 times stronger than the process of within-district selection across communities in order for the observed $\hat{\tau }^{Gov't}$ to be driven entirely by selection on unobservables. While a correlation between NDC vote share and the degree of endogenous selection of projects into fund sources cannot be ruled out, it seems implausible that selection on unobservables accounts for more than a tiny share of the differential effect of NDC vote share on government-funded projects relative to donor-funded ones.

TABLE 3.

Community Fixed Effects

Note: Dependent variable is project completion. Sample is all projects with nonmissing community name. All specifications include project type and characteristics as in Table 2. Year fixed effects included in Column 2. Huber-White robust standard errors clustered by district-year in Column 1 and community in Column 2.