I. INTRODUCTION

Over the last 20 years geopolitical and global economic developments have modified the way official foreign aid is given. The so-called new rhetoric on aid has recognized the importance of encouraging greater ownership of development programs in recipient countries (e.g., see the Paris Declaration on Aid Effectiveness, OECD 2005). In particular, ownership has been seen as crucial for the successful implementation of conditional reform programs and basing reform designs on context-specific knowledge could be one way to stimulate recipient countries' ownership. (1)

Donors that aim at maximizing ownership could be expected to grant substantial leeway to the recipients of their aid. Donor and recipient preferences on how to use aid can however differ. Donors use parts of their aid to promote development and improve policies and institutions (Fleck and Kilby 2010), while recipients might want to use it to grant political favors to their preferred constituencies or delay the implementation of reforms (Dreher et al. 2014; von Borzyskowski 2016). These differences in preferences give donor countries incentives to keep control of how recipients spend the aid. The differences in preferences about how to use aid imply a trade-off when deciding about whether and to what extent control over the aid should be given to the recipient or kept with the donor. The trade-off is complicated by the role that donor and recipient country information--and the way such information is exchanged between them--plays in how to make best use of the aid with respect to developmental outcomes. The link between the difference in preferences of donors and recipients in how to use aid, the relative importance of donor and recipient information, and whether and to what extent this information is communicated between them, is the focus of this paper. (2)

Countries' local knowledge often consists of unverifiable information (or verifiable only at a cost) and so the quality of the information the recipients provide to the donors crucially depends on the conflict of interest between the recipient (the sender of the information) and the donor (the receiver). Communication is complicated by the fact that donors also own some private information that is relevant to the implementation of effective polices. In this setting, mutual communication is important as the donor possesses skills and information which are useful in processing the country's local information. Thus, a combination of the private information of the donor with those of the recipient is required for the design of the "optimal" policy package. The analytical setting is one of two-sided incomplete information where agency problems have the indirect negative effect of preventing full communication between the sender and the receiver.

As in the study by Marchesi, Sabani, and Dreher (2011), we model the transmission of information between the donor and the recipient country as a cheap talk game (Crawford and Sobel 1982). Information is assumed to be "soft" and the transmission of information to be costless. We compare two types of incentive schemes (delegation vs. centralization) relative to the quality of the transmitted information. We define "centralization" as a framework in which control rights over policies are assigned to the donor. On the contrary, we define "delegation"--or "decentralization"--as a framework in which the recipient country is left with considerable freedom to devise its own policy actions.

We consider a situation in which the recipient is biased in favor of the "status quo," whereas the donor is biased in favor of more (or deeper) policy reforms relative to the recipient. What we have in mind here is a situation where recipient governments might be corrupt and incompetent, maintain unsustainable economic policies like high inflation and budget deficits, or repress minorities. We assume donors want to use their aid to achieve changes to the status quo, but face resistance by the recipient government. With both delegation and centralization, such misalignment of interests prevents full communication. Therefore, the optimal allocation of control rights over policies from the donor's perspective will depend on the relative importance of the two parties' information. It will also depend on the degree to which donor and recipient preferences differ (which we refer to as the "agency bias"), simultaneously affecting the amount of information transmitted and the degree of reforms implemented.

In line with Marchesi, Sabani, and Dreher (2011), the main theoretical findings are as follows. For a given agency bias, when the recipient's local knowledge is more important than the donor's information, their discretion in the choice of policies (delegation) should be increased. Conversely, there should be less freedom in designing policies (centralization) when the donor's information is more relevant. As far as the effect of the agency bias is concerned, there are two opposing effects. Intuition would suggest that an increase in the conflict of interest between the donor and the recipient would make the donor more inclined toward "centralization." The agency bias, however, also affects the quality of communication and--since an increase in the bias reduces the amount of information transferred to the donor by the recipient--the donor's incentive to delegate may increase, particularly when local knowledge is crucial for designing the donor's preferred policies.

An immediate empirical implication of the model is to investigate the way in which aid is committed in relation to information transmission problems. We focus on two distinct ways of delivering aid, budget support and project aid. Budget support increases the involvement of the recipient governments in the decision-making process and is thus an example of a "delegation-scheme." This is because budget aid is directly transferred to the recipient government and can be used by the recipient at some discretion. (3) Conversely, project aid represents a more "centralized" type of aid. Donors and recipients negotiate the specific projects the aid is given for. What is more, donors are usually involved in the details of preparing and implementing the project, leaving little discretion on how to use the aid. We therefore consider the relative importance of donor and recipient private information--and the difference in their preferences on how to use the aid--as determinants of project and budget aid.

We test our theory focusing on aid given by the 28 bilateral donors of the OECD's Development Assistance Committee (DAC) to a maximum of 112 recipient countries over the 1995-2010 period, resulting in more than 45,000 observations at the donor-recipient-year level. We measure the bias in donor and recipient preferences with a number of proxies, among them a measure based on how they vote in the United Nations General Assembly (UNGA) on a broad range of topics. Our proxies for the availability of information to the donor relate to how transparent recipient country policies, data, and local environments are for the donor. Controlling for the main donor- and recipient-country variables that determine the dyadic aid relationship, and donor-recipient-pair- as well as yearfixed effects, we find that misaligned interests and informational asymmetries differentially influence whether donors grant their aid as project aid or budget aid, in line with our theory.

The article is organized as follows. Section II briefly describes the related literature. A sketch of the model is developed in Section III. Section IV introduces our data, while Section V describes the empirical model and our results. Section VI concludes.

II. RELATED LITERATURE

This study relates to two strands of literature. The first is the literature on aid allocation and selectivity. This literature tries to disentangle the various motives of donors when giving aid, usually referring to commercial, geo-strategic, developmental, and "good policy" related motives (see, e.g., Alesina and Dollar 2000; Dreher, Sturm, and Vreeland 2009a, 2009b; Kuziemko and Werker 2006). Most directly related to the question we focus on in this paper are studies that address the choice between project and budget aid (see, e.g., Chauvet, Collier, and Fuster 2013; Clist, Isopi, and Morrissey 2012; Cordelia and delTAriccia 2007; Hefeker 2006; Koeberle, Stavreski, and Walliser 2006; Morrissey 2006; Mosley and Abrar 2006; Ouattara and Strobl 2008).

Cordelia and delTAriccia (2007) relate the choice between project aid and conditional budget support to the different preferences between donors and recipients. They show that budget support is preferable to project aid when the donor's preferences are close to those of the recipient and the amount of aid is small relative to the recipients' own resources. Morrissey (2006) also finds that budget support can safely be granted if recipients allocate spending broadly as agreed with their donors. Rather than imposing prior actions on the recipient, donors should then focus on the effectiveness of such spending when determining eligibility to budget support. (4) In a similar vein, Mosley and Abrar (2006) show that trustful relations between donors and recipients are fundamental for the effectiveness of conditionality, and in particular for those of budget support.

More recently, Chauvet, Collier, and Fuster (2013) have also related the existence of a conflict of interest between donors and recipients to the choice of (donors' supervision of) project aid. Applying principal-agent theory to the performance of aid projects they show that in a wide range of circumstances the donor should put greater effort into supervision when the difference between the agent's preferences and its own is greater. They test this prediction using data on World Bank project performance and--consistent with their theory--find that donor supervision of projects is significantly more effective in improving project performance when interests are widely divergent. Like we do in this paper, Cordelia and delTAriccia (2007) and Chauvet, Collier, and Fuster (2013) use a principal-agent framework and relate the conflict of interest between donors and recipients to the choice of whether to give aid as budget support or project aid. None of them, however, has considered the importance of communication between the donor and the recipient for the design of policies, nor have they related the choice between these different aid schemes to the importance of fostering communication between the two.

The second strand of literature to which this paper relates is primarily concerned with the role of donors (or lenders) in designing development reforms and thus to the importance of enhancing recipients' ownership. (5) The principle that ownership is crucial for the (successful) implementation of reforms is now well established. As emphasized by various studies including Easterly (2008), Dixit (2009), Besley and Persson (2011), and Marchesi, Sabani, and Dreher (2011), institutions and policies are context-specific and donors and lenders do well to base their policies on a good knowledge of the recipient country's characteristics, which in turn implies greater ownership of policies in recipient countries. (6) Nevertheless, the mechanisms and circumstances under which such knowledge should be transferred have rarely been investigated. (7)

Finally, we should mention that the use of dyadic data is crucial in our setting. While much of the aid allocation literature uses monadic data, (8) our model's implications for donor-recipient communication requires measures of donor-recipient-specific relationships (such as the political or ideological distance between them and bilateral trade). Such data allow us to take into account the salience of the informational asymmetry (and of the agency bias) in donor-recipient pairs in addition to more general indicators of recipient transparency (or bias). For example, even in generally intrans-parent recipients, for which the use of local information should in principle be important, specific donors might still decide to opt for centralization, for example because they have some country-specific knowledge due to a prolonged relationship with a particular recipient (which has reduced the relative importance of the recipient's knowledge over time).

We contribute to the literature both theoretically and empirically. Regarding theory, we analyze the transmission of information in the allocation of aid. To our knowledge, it is the first time that communication is explicitly introduced to the context of foreign aid. With respect to our empirical models, even though some papers have considered the importance of distinguishing among different types of aid flows and some have empirically investigated the determinants of budget support, we are the first to test whether this choice is responsive to communication between the donor and the recipient.

III. THEORETICAL FRAMEWORK

The framework relies on the model of Marchesi, Sabani, and Dreher (2011), which we modify in order to be applicable to the issues central to this paper. The main change with respect to Marchesi et al. (as well as to Flarris and Raviv 2005) relates to the different environment in which we investigate the cheap talk.

To analyze whether the donor has an incentive to delegate the control of decision-making to recipient governments we focus on the aspects of the model that are central to derive our hypotheses. For reasons of clarity, all detailed derivations and proofs are however shown in Appendices S1 - S3, Supporting Information. (9) The model features two players--the donor and recipient countries' governments--that own different types of information both required for the optimal choice of policies in the recipient country (in the context of disbursing aid), denoted by p. The recipient country's welfare is proxied by Y(p) (i.e., the country's per capita national income), which is a function of policy p. The policy maximizing Y(p) is denoted by [p.sup.*]. In turn, optimal policy is defined by [p.sup.*] = g + d, where g and d are stochastic variables that proxy for information observed only by the recipient government and, respectively, the donor government; g and d are independently and uniformly distributed on the intervals [0, G] and [0,D], respectively. This captures that the larger the interval [0, G] ([0, D]), the larger the informational advantage of the recipient (donor).

The recipient's superior information over g represents the local knowledge (for instance information about the country's economy and sociopolitical characteristics or better knowledge about the risks and opportunities of local investment projects), which can be seen as deriving from its closer proximity to the country's culture and business environment as compared to the donor. The recipient's informational advantage may depend not only on how relevant its knowledge is per se, but also on how valuable such information is relative to the donor. For example, in highly intransparent environments such informational advantages would be more salient compared to more transparent ones.

However, the donor's informational advantage d is derived from its cross-country knowledge. For example, a donor that has implemented projects in the health sector in a number of different countries has accumulated project-related knowledge that will be valuable for the implementation of health projects in the recipient country. Both types of information are assumed to be (at least partly) "soft," that is, they cannot easily be certified.

Events unfold in three stages: allocation of control rights by the donor, communication, and policy implementation. (10) In the first stage, the donor either allocates authority over the choice of the policy vector to the recipient government or retains authority. Centralization refers to the scheme in which the donor decides on the policy vector, whereas under decentralization control rights are allocated to the recipient government. After the first stage of the game, the real state of the world is revealed to both players. In the second stage, communication takes place. Under centralization, the government sends a "message" to the donor regarding its "local knowledge." Upon receiving the message, the donor updates its beliefs and chooses the policy vector. Under decentralization, the donor sends a message to the recipient concerning its private knowledge of the state of the world. In this case, the government updates its beliefs and chooses the policy vector. Finally, in the third stage, the policy is implemented and outcomes are realized.

The donor maximizes the following objective function:

(1) [U.sup.D] = [U.sup.D.sub.0] - [(p - [P.sup.*.sub.D]).sup.2],

where [U.sup.D] decreases with the distance between the actually implemented policy p and the optimal policy [P.sup.*.sub.D], with [U.sup.D.sub.0] = [U.sup.D] ([P.sup.*.sub.D]). (11) The optimal policy choice of the donor deviates from the optimal policy [p.sup.*] by a factor e > 0 (i.e., [P.sup.*.sub.D] = [p.sup.*] + e). Besides the recipient country's income, the donor cares as well for the externality that the choice of a specific policy p could have on the donor country's economy (e.g., regarding access to markets or natural resources), in line with the aid allocation literature (e.g., Alesina and Dollar 2000; Dreher, Nunnenkamp, and Thiele 2011; Kuziemko and Werker 2006). e > 0 captures the extent to which the policy choice of the donor may deviate from its optimal level [p.sup.*] due to the pressure of such distortions.

The recipient country's government maximizes:

(2) [U.sup.G] = [U.sup.G.sub.0] - [(p - [P.sup.*.sub.G]).sup.2],

which is decreasing in the distance between the implemented policy p, and the recipient government's preferred policy [P.sup.*.sub.G], with [U.sup.G.sub.0] = [U.sup.G] ([P.sup.*.sub.G]). (12) The optimal policy choice of the government deviates from the optimal policy [p.sup.*] by a factor b > 0 (i.e., [P.sup.*.sub.G] = [p.sup.*] - b). The recipient government cares about its national per capita income, but its choice may be constrained by the influence of some interest groups benefitting from structural distortions (e.g., Drazen 2002). b > 0 captures the extent to which the policy choice of the recipient may deviate from its optimal level [p.sup.*], for example due to the pressure of interest groups opposing policy reforms (among others, Alesina and Drazen 1991; Fernandez and Rodrik 1991; Tabellini and Alesina 1990). (13)

Therefore, the difference in optimal policies is given by

(3) [P.sup.*.sub.D] - [P.sup.*.sub.G] = [P.sup.*] + e - ([P.sup.*] - b) = e + b = B,

where B reflects the extent of any conflict of interest among donors and recipients over the desired policies that might lead to a deviation from the first best policy [p.sup.*], including, but not limited to, the presence of some geo-political distortions at the donor level, and the pressure of local interest groups and re-election concerns.

A. Communication Game

The donor can choose between centralization or delegation. Opting for centralization, the donor minimizes the costs of misaligned incentives and makes full use of its private knowledge. At the same time, it underutilizes the recipient's information. Under delegation, the donor allocates policy decision-making to the recipient. While in this case the recipient's private knowledge is fully exploited, the results can deviate from the donor's optimal policy (loss of control).

In the communication equilibrium, the recipient government only learns the interval to which the realization of d belongs, and hence obtains only incomplete information about the donor's knowledge. The smaller the size of the partition interval, the more informative the donor's message. We denote the maximum number of intervals--N(D,B)--as a function of the bias B and the length of the partition of the donors' knowledge D. Following Crawford and Sobel (1982), the most informative equilibrium--in which the number of intervals N is maximal--always exists and is a focal equilibrium of the communication game.

In the focal equilibrium, the donor's ex ante expected welfare loss increases with the importance of the donor's private information D, since the donor's private information is not fully exploited under delegation. Finally, for any given D, the maximum precision of the information transmitted by the donor decreases with the extent of the bias B (i.e., the larger the bias B, the less precise and informative cheap talk will be). However, if the donor chooses centralization, it fully exploits its own information D and chooses its preferred policy vector [P.sup.*.sub.D] As centralization results in an underutilization of the recipient's information G, the donor's ex ante expected loss is increasing with the recipient's informational advantage.

The donor determines whether or not to retain its control rights over policies by comparing its ex ante expected loss under delegation with its expected loss under centralization. Since both are increasing in D (under delegation) and G (under centralization), we can identify cut-off values of D and G at which the scheme choice switches. The scheme choice, thus, depends on the extent of the conflict of interest (B) and the relative importance of the two players' respective informational advantage (D, G).

Figure 1 represents the choice between centralization and delegation as a function of D and G. The threshold D(G,B) is upward sloping, and divides the (G,D) plane into two regions (centralization and delegation) lying below the 45[degrees] line. The donor will opt for delegation only if the recipient's private information G is (strictly) greater than its own private information D and greater than the threshold level D(G,B). The delegation region is smaller than the centralization region: the agency bias B requires G to be strictly greater than D in order for delegation to be optimal. This holds because the loss due to underutilization of the recipient's information is compensated for by the elimination of the bias and the full exploitation of the donor's own private information D. Conversely, the donor always chooses centralization whenever its private information D is more important than the recipient's private information (i.e., D > G). Additionally, it opts for centralization if D(G,B) [less than or equal to] D < G, that is, even when the recipient's informational advantage G is greater than D, but smaller than the threshold value D(G, B). (14)

In general, as Figure 1 shows, the threshold D(G,B) is not monotone in the bias B, as an increase in B has both direct and indirect effects. Directly, an increase in B increases the agency problem, thus reducing the donor's incentive to delegate. Indirectly, an increase in B also reduces the equilibrium amount of information transferred by the recipient to the donor under centralization, thus making delegation a better choice. Therefore, an increase in the agency bias, while making the recipient's choice less attractive to the donor, can also decrease the incentives of the recipient to communicate its private information in the centralization game more than in the delegation game. The net effect can result in switching from centralization to delegation with an increase in the bias, in order to make full use of the recipient's private knowledge. (15)

B. Empirical Implications

The model provides some normative indications regarding the allocation of control rights over policy actions in the donor-recipient relationship, and testable implications can be derived from the theory. The main prediction of the model is that delegation should prevail when the "loss of information" dominates the "loss of control." This holds true when the importance of the recipient's knowledge--to be partially lost under centralization--dominates the importance of the donor's private information, for a given agency bias. To the contrary, centralization should prevail when, for a given bias, the importance of the donors' knowledge dominates the role of the recipient's local knowledge.

Since budget support increases the involvement of recipient governments in the decision-making process, it is an example of "delegation" in the sense of our model. Aid in the form of budget support is directly given to the recipient, so that control over the aid money rests with the government of the recipient country rather than the donor. (16) Alternatively, project aid represents a good example for a more centralized provision of aid. Projects are usually selected in close collaboration with the donor, and are closely supervised, or even directly implemented by the donor, thus leaving less influence for the recipient government. (17)

We empirically investigate whether or not the share of project aid and budget aid (to overall aid commitments) are affected by variables related to the relative importance of donor-recipient informational asymmetry and by variables capturing the size of the agency bias, holding recipient country characteristics, their economic performance, and the dyadic relation between the donors and recipients (as well as donors' political motivations) constant. Specifically, for any given bias, budget support (or delegation) should be preferred in countries whose local knowledge is relatively more important. Conversely, project aid (or centralization) should prevail when the recipient's local information is less crucial.

A second important feature of the model is the presence of a nonmonotonic relationship between delegation and the misalignment of interests between the donor and recipient. The bias has both direct and indirect effects working in opposite directions. The direct effect is to increase the agency problem, thus reducing the donor's incentive to delegate. The indirect effect both reduces the amount of information transferred by the donor to the recipient under delegation (leading to centralization) while at the same time reducing the amount of information transferred by the recipient to the donor under centralization (leading to delegation). The overall effect of the agency bias on delegation cannot be analytically derived. However, we know that the indirect effect works through the importance of information.

The donor's informational advantage may depend not only on the relevance of its knowledge per se, but also on how valuable such information is relative to those of the donor. In less transparent countries informational advantages are arguably more salient as compared to more transparent ones. Therefore, we expect that the indirect effect of the bias on delegation will prevail when the information transferred by the recipient is of higher value to the donor--that is, in highly intransparent environments. However, we expect the direct effect to prevail when the information transferred by the recipient is relatively less important to the donor, namely in more transparent environments.

Greater transparency increases the share of "hard" information that can easily be transferred and decreases the importance of private "soft" knowledge, hence making the informational asymmetry less salient. (18) As a consequence, given the trade-off between loss of control and loss of information faced by the donor, in order to disentangle the direct and the indirect effects of the bias, we interact the "bias" with "transparency." To the extent that aid is allocated in line with the model, we expect to find a negative (or insignificant) interaction between the two, as the importance of the local information for the donor decreases with transparency, reducing the (indirect) effect of the bias on decentralization. The easier donors can access specific local knowledge, the lower the likelihood that they are willing to delegate decision-making authority based on the importance of this knowledge. Budget aid therefore becomes less, and project aid more likely.

IV. DATA

We examine the determinants of budget aid and project aid in a dyadic donor-recipient setting. Data on general budget support and project aid are from the DAC's Creditor Reporting System (CRS; OECD 2016). These data are not reported for years prior to 1995. We estimate separate regressions for the two types of aid rather than using the ratio of the two, which could reflect their relative importance in one regression. The reason is that many countries receive no aid from a particular donor, while others receive aid of only one of the two types. (19) Zero aid could then not be separated from aid of the category we would put on the numerator; zero aid in the denominator would make the share approach infinity. We avoid both problems by investigating the two types of aid in separate regressions, and comparing the relative influence of our variables of interest in determining the shares of these flows in overall aid commitments. (20) In the following, we propose a number of proxies to measure the extent of the agency bias and the relative informational advantages of the donor and recipient governments.

A. Variables of Interest

Our variables of interest are meant to capture the extent of the agency bias, the donor country's general knowledge, and the recipient country's local knowledge--and how easily this information is available. We introduce them in turn.

Agency Bias. Empirically, we are interested in a bias in the objective function of the recipient country's authorities relative to the preferences of the donor. According to the political economy literature, measures of political instability, polarization, and social division (e.g., Alesina and Drazen 1991; Tabellini and Alesina 1990) account for a country's "resistance" against reforms (or status quo bias). (21) For any given policy environment, such countries will find it more difficult to make changes to their policies to reduce a given agency bias. With this in mind, we included proxies for Government Capability and Ethnic Tensions, taken from the International Country Risk Guide (ICRG). Government Capability ranges from 0 to 12 and is "a measure of the government's capability in carrying out its declared programs/policies and its ability to stay in office." Higher values on this measure of institutional capacity imply that it is easier for the government to overcome internal resistance and implement reforms. The status quo bias is thus lower. We include Ethnic Tensions, which ranges from zero to six, with higher values indicating more tensions. Ethnic Tensions measure "the degree of tension within a country which can be attributable to racial, cultural and language divisions" (PRS Group 1998). (22) It is thus a proxy for polarization and social division, which the previous literature has shown to inhibit reforms. At the same time, highly polarized countries tend to exhibit a certain degree of "favoritism" in their preferred policies (Franck and Rainer 2012), arguably in opposition to the average donor's preferences. (23)

In order to capture the dyadic component of the agency bias between specific donor-recipient-pairs, we include UNGA voting alignment (UNGA Distance), which captures the political distance between the donor and recipient. Specifically, we include the dyadic distance between their ideal points in the UNGA, estimated using a dynamic ordinal spatial model (Bailey, Strezhnev, and Voeten 2017). This ideal point distance is constructed to measure the government's preferences on foreign policy, and to be comparable over time. It is based on dyadic voting data in the UNGA, but refers to a single dimension in order to avoid measuring shifts in countries' policy positions when in fact only the composition of topics in the UNGA has changed, but preferences on each individual topic have not. While variation in UNGA voting is most substantial across countries, there is considerable variation within countries over time as well (Dreher and Jensen 2013; Rommel and Schaudt 2016). UNGA Distance is thus well suited to measure time-varying differences in preferences over policy. (24)

According to our model, the effect of the agency bias on the extent of reforms can go either way. As we described above, an increase in the agency bias per se has both direct and indirect effects which could either reduce or increase the incentive to delegate, depending on which of the two effects dominates, on average.

Information. We expect the importance of a recipient's local knowledge to increase with the salience of the informational asymmetry between donors and recipients. In particular, local knowledge is crucial for intransparent countries as less transparency decreases the share of "hard" knowledge and increases the importance of "soft" knowledge to be obtained from the recipient. In this context, facing the trade-off between loss of control and loss of information, donors might decide to give more importance to information and, in turn, give greater control to the recipient. If this is the case, donors will prefer budget aid to project aid. In order to measure the importance of a recipient's local knowledge, our main index follows Hollyer, Rosendorff, and Vree-land (2011), who suggest missing data on standard economic indicators (relating to economic policy and debt) as indicators of (lack of) transparency. As Hollyer, Rosendorff, and Vreeland (2011, 1198) point out, this "measure of transparency [...] directly reflects government decisions to release accurate economic data." Rather than choosing any arbitrary data series we evaluate all 1260 series included in the World Bank's (2013) World Development Indicators. Our resulting indicator for Transparency shows the share of series for which there are data available in a given country and year. (25) As we therefore treat missing data as (lack of) information, the resulting indicator has the advantage that it is available for all countries and years.

An additional proxy for the availability of information is the number of Telephone Lines (per 100 people), which is also widely available. As we explain in Dreher et al. (2016), this variable can be seen to proxy for all kinds of technological barriers to information transmission. Which technology is most relevant to capture information transmission varies over time, so that the easy availability of internet access or mobile phones will better proxy for information transmission in more recent years, but not in the earlier years of our sample. According to the results by Chung, Fleming, and Fleming (2013), Telephone Lines exerted the strongest effect on trade among a number of alternative proxies for the quality of information and communication technology. As Dreher et al. (2016) point out, the number of Telephone Lines is highly correlated with a combined media access variable ([rho] = 0.80) and a variable capturing the number of computers per capita ([rho] = 0.87) in those periods where both are available. (26)

An additional way of measuring the salience of the informational asymmetry is by including information on the dyadic relationship between specific donors and recipients. We therefore construct a measure for Bilateral Experience, calculated by the number of years since a donor has first given aid to a specific recipient country. (27) When countries have a longer bilateral aid relationship the recipient's local knowledge seems less important compared to the donor's knowledge. This is because the donor has gathered experience through previous aid projects and is thus better informed than without having this country-specific experience, on average. The need for delegation is therefore reduced by the number of years since the donor had first committed aid to the recipient. Similar to Bilateral Experience, Bilateral Trade proxies for dyadic donor-specific information about the recipient country. The importance of information costs in determining trade is well established (e.g., Fink, Mattoo, and Neagu 2005). While causality between information and trade can be either way, we thus interpret Bilateral Trade as an additional informational variable at the dyadic level.

Poor quality of recipient government staff could also be a reason for a recipient to seek a donor's technical advice and could thus explain the choice of project aid over budget aid. In order to capture the quality of recipient government staff, we include the ICRG index of Bureaucratic Quality. Bureaucratic Quality ranges from zero to four, with higher values showing "better" environments. High scores in Bureaucratic Quality indicate that the bureaucracy has the strength and expertise to govern, without the necessity for drastic changes in policy or interruptions in government services.

Finally, as a measure that is specific to the donor, rather than the recipient, we calculate the number of recipients a donor gives aid to in a particular year to proxy for the donor's most recent cross-country knowledge. The number of recipients a donor gives aid to at the same time proxies for the donor's information about development policies implemented in different countries and contexts and the global environment in which these policies are embedded at a particular point in time (Donor Experience).

B. Control Variables

Much of the literature on aid allocation has evaluated whether and to what extent commercial and political donor interests have shaped the allocation of aid, but recipient country "need" and "merit" have also featured prominently (Claessens, Cassimon, and van Campenhout 2009; Dollar and Levin 2006; Fleck and Kilby 2010; Hbffler and Outram 2011). Our main specification is parsimonious, controlling for (log) GDP per capita to take account of development, and (the log of) Population which also captures "need," but can as well be taken as proxy for the ease of obtaining a country's political cooperation (as smaller countries are easier to "buy"; see, e.g., Boone 1996), and is thus a proxy for the donors' political interests. (28)

We provide the details of the definitions and sources of the variables included in the regressions and descriptive statistics in Appendices S4 and S5. Appendix S6 shows the correlations of the variables included in the analysis.

V. METHOD AND RESULTS

We use data for the 1995-2010 period and a maximum of 112 recipient countries, due to data availability. The dependent variables are defined as shares of total dyadic aid commitments. The regressions are estimated with ordinary least squares (OLS), at the donor-recipient year level. We estimate the model with country-pair-fixed effects and year-fixed effects (clustering standard errors at the country-pair-level). (29) The regression equations are:

(4) [P.sub.i,j,t] = [[beta].sub.1][X.sub.1i,t-1] + [[beta].sub.2][X.sub.2i,j,t-1] + [[eta].sub.i,j] + [[tau].sub.t] + [u.sub.i,j,t],

and

(5) [B.sub.i,j,t] = [[beta].sub.1][X.sub.1i,t-1] + [[beta].sub.2][X.sub.2i,j,t-1] + [[eta].sub.i,j] + [[tau].sub.t] + [u.sub.i,j,t],

where [P.sub.i,j,t] and [B.sub.i,j,t] represent project and budget aid commitments as a share of overall commitments from donor j to recipient i in year t, and [X.sub.1] and [X.sub.2] are vectors containing the variables introduced above (lagged by one year). While [X.sub.1] is the vector of recipient-specific variables, [X.sub.2] includes variables that vary over donor-recipient-pairs. In one set of regressions [X.sub.1] and [X.sub.2] include interactions between Transparency and our dyadic proxy for the agency bias, allowing us to disentangle the average effect of the bias according to whether transparency is high or low. Finally, [[eta].sub.i,j] and [[tau].sub.t], are donor-recipient-pair--and year-fixed effects, respectively, while [u.sub.i,j,t] is the error term.

Contrary to most of the aid allocation literature, (30) we estimate rather conservative models, which include country-pair- and year-fixed effects, and lag the explanatory variables by one year. We therefore control for unobserved effects that exclusively vary at the country-pair- and year-level, substantially reducing concerns over endogeneity. What is more, we investigate aid provided by all 28 bilateral DAC donors rather than aid from a particular donor. We can therefore account for a variety of observable indicators at the recipient- and donor-level as well as on the donor-recipient-pair-level, including information on historical, political and economic ties. While this does not provide a bullet-proof identification strategy, we are more conservative than most of the related literature. (31) Still, we prefer to interpret the coefficients in the models below as conditional correlations rather than causal effects.

We report the basic results in Table 1. Column 1 shows the results for project aid excluding the dyadic variables, while column 2 shows those for budget aid instead. Columns 3 to 6 include characteristics of the country-pair--the dyadic transparency indicators in columns 3 and 4 and UNGA Distance in columns 5 and 6. Odd column numbers focus on project aid, while even column numbers refer to budget aid (both measured as a share of overall dyadic commitments). Across regressions, the share of budget aid increases with GDP per capita and Population, while there are no consistent correlations of these control variables with the share of project aid, at conventional levels of significance.

As can be seen in Table 1, the results are in line with our hypotheses regarding the effect of what we call "informational variables," on the provision of aid. In all regressions, budget aid and project aid increase with greater Transparency. The coefficients for project aid are however larger compared to those of budget aid in all regressions, indicating that donors prefer a type of aid that allows them to keep control when it is comparably easier for them to access recipient information. (32) Project aid--but not budget aid--increases with the availability of Telephone Lines and greater Donor Experience as well, indicating the importance of the informational infrastructure for donors' preference of project aid over budget aid. Quantitatively, an increase in Transparency by 0.1 (the mean being 0.65 in column 1) increases the share of project aid in overall aid commitments by between 0.019 and 0.035 percentage points. An increase in Transparency by one standard deviation (i.e., 0.12) increases the share of project aid in overall aid commitments by between 0.02 and 0.04 percentage points. For the average recipient country in our sample, this represents a six to 12% increase. A one standard deviation increase in Telephone Lines (representing 13 telephone lines per 100 people) and a one standard deviation increase in Donor Experience (i.e., 47 recipient-years) imply an increase in the share of project aid in overall commitments by between 7 and 10 percentage points, respectively.

The results also show that donors prefer project aid over budget aid with a longer bilateral aid history in the recipient country, indicating that donors who are less in need of recipient information delegate less. The longer a donor has been giving aid to a recipient, the more experience and knowledge it has accumulated. Consequently, the informational advantage of the recipient is reduced, which leads to a positive correlation with project aid but not budget aid. Specifically, one more year of Bilateral Experience leads to an increase of 0.015 percentage points of overall aid commitments, which represents a yearly increase of 4%-5% for the average recipient country. In a similar vein, we introduced Bilateral Trade as an additional proxy of dyadic information. Again, the results are as expected: Bilateral Trade significantly reduces the amount of budget aid, in line with the prediction of the model regarding the importance of information for the choice of delegating aid policies.

As we are interested in the statistical significances of groups of variables rather than of individual variables, we rely on tests of their joint significance in order to evaluate our hypotheses. Specifically, we evaluate the relative importance of the transparency variables based on their joint significance in the budget aid--compared to the project aid--regressions. An F test indicates that all transparency variables are jointly highly significant for project aid (column 5), but only marginally significant for budget aid (column 6).

While we consider tests for the joint significance of our variables of interest to be most appropriate to test our hypotheses, note that most of the variables are also individually significant. The exception is Bureaucratic Quality, which is completely insignificant in all regressions, with no significant differences between project aid and budget aid. The correlation between budget aid and Transparency is significant at the 1% level in column 1, but much weaker in significance when we add the dyadic proxies for transparency, as could be expected. (33) The correlation between Telephone Lines and project aid, as well as those between Donor Experience and project aid, is significant at the 1% level in all regressions though (and insignificant for budget aid). Bilateral Experience is significant at the 1 % level for project aid, but weakly significant or insignificant for budget aid, while Bilateral Trade is (negatively) significant at the 1% level for budget aid only.

In summary, we find strong evidence that donors allocate their aid in line with the "transparency"-related predictions of our model. Since transparency indicates the relative importance of the recipient's knowledge (as compared to the donor's knowledge), more transparent countries receive more project but not budget aid, as our theory implies. Donor countries do not need to rely on the recipient's local knowledge if transparency is high.

Our model is less clear-cut when it comes to making predictions about the differences in donor and recipient preferences ("bias"). As we have outlined above, the effect of the bias on delegation could be either direct (reducing delegation) or indirect (increasing delegation by reducing the amount of communication under centralization). We thus do not have strong predictions for the effect of the "bias-related" variables, on average.

According to the results in Table 1, the recipient-specific measures for the agency bias do not turn out to be significant determinants of the choice between project and budget aid, on average. We find no effect of the recipient country's Ethnic Tensions and Government Capability. We should point out that Ethnic Tensions are related to the importance of both the bias and the recipient's private information (when a country is more multifaceted from a social point of view its local knowledge is more important). As a result, since these two effects could go into two opposite directions, the insignificant coefficient of Ethnic Tensions is easy to explain. In the case of Government Capability, our theory suggests that lower government stability should lead to higher (lower) amounts of project (budget) aid, as one would expect centralization to be higher in more unstable countries. This argument, however, overlooks the fact that an increase in the bias also has the effect of reducing the degree of communication under centralization, thus making such a decision costlier. The net effect could as well then lead to an insignificant coefficient. (34)

We find significant coefficients for our dyadic proxy UNGA Distance. Specifically, while the share of budget aid decreases with UNGA Distance, project aid is unaffected. Centralization thus dominates delegation when the bias of the recipient country relative to the donor is too large according to this dyadic measure. (35) The direct effect (reducing delegation) thus dominates the indirect effect (increasing delegation), on average.

In order to disentangle the direct and indirect effects of the bias on delegation, we investigate how transparency and agency bias interact. We focus on the interaction between UNGA Distance and the level of Transparency. Table 2 presents the results on the differential effect of the agency bias (UNGA Distance) conditional on the level of Transparency and Telephone Lines.

Columns 1 and 2 interact Transparency with UNGA Distance. As can be seen, the effect of UNGA Distance on project aid decreases with Transparency (column 1), while its effect on budget aid becomes stronger (column 2), at least at the 10% level of significance. Results are similar but statistically weaker when we turn to the interaction of Telephone Lines with UNGA Distance in columns 3 and 4. While the direct or indirect effect could dominate according to our model on average, the indirect effect of the bias should prevail in highly intransparent environments, where the information transferred by the recipient is of higher value to the donor. As we are interested in how the marginal effect of the agency bias changes over the range of the transparency indicators, we calculate average marginal effects and show them in Figures 2 (project aid) and 3 (budget aid), in tandem with 90% confidence intervals.

Figures 2 and 3 show that the marginal effect of UNGA Distance on the amount of project aid decreases with the intensity of Transparency, while its effect on the amount of budget aid increases with Transparency. Both effects are significant for low levels of Transparency, but turn insignificant at conventional levels when Transparency is high. As transparency increases, the recipient's local knowledge becomes less relevant, so that donors prefer centralization (project aid) to delegation (budget support). Only when transparency is high, the size of the bias loses relevance in predicting the difference between project aid and budget aid. In highly transparent countries, donor countries do not depend on recipient government information and so depend on communication to a lower extent. Overall, these patterns fit our model's predictions well. (36)

VI. CONCLUSIONS

In this paper we have explored the role of information transmission between a donor and a recipient country in explaining how donors allocate budget aid and project aid. By relating the quality of the information supplied by a recipient country to the donor (and vice versa) to the misalignment of interests between the two, we analyzed the properties of different aid schemes relative to the quality of the transmitted information. More specifically, we have compared an aid scheme in which control rights over policies are allocated to the donor, that is, centralization (or project aid), with an aid scheme in which the recipient is left with more freedom to devise its own policy actions, that is, delegation (or budget support).

The main theoretical findings are as follows. For a given agency bias, when recipients' local knowledge is more important than the donors' information, their discretion in the choice of reforms (delegation) should be increased. Conversely, there should be less freedom in designing reforms (centralization) when the donors' information is more relevant. The impact of the agency bias on determining the optimal lending scheme remains a priori undetermined as it can have two countervailing effects at the same time (a direct and an indirect one).

In the empirical section, we focused on two distinct ways of delivering aid, budget support and project aid. Budget support increases the involvement of the recipient government in the decision-making process and is thus an example of delegation. Conversely, project aid represents a more centralized type of aid. We investigated the role of the relative importance of donor and recipient information in determining which aid scheme is preferred. Controlling for countries' characteristics, their economic performance and dyadic relations between donors and recipients, we find that transparency does influence the relative amount of project versus budget aid. More specifically, as transparency increases, donors prefer project aid to budget support. As the agency bias is concerned, the results of our dyadic measure are in line with our theoretical predictions, according to which centralization should dominate delegation when the bias is too large. Finally, the marginal effects of the bias, conditional on transparency, point to the dominance of the direct over the indirect effect when transparency is low, leading to a centralization scheme.

Our model suggests that donors who allocate aid taking properly account of information and preferences will achieve the results they aim for more effectively. Whether an allocation of aid in line with the model is likely to increase economic growth or reduce poverty depends on whether the donor is sufficiently benevolent. According to parts of the aid effectiveness literature, however, both project aid and budget aid have not on average been effective with respect to achieving growth (e.g., Rajan and Subramanian 2008). This could imply that donors in reality allocate aid in line with other, geopolitical or commercial, targets. (37) It could also imply that the targeting of aid toward budget or project aid, while significant, is not yet sufficiently elaborated. To the extent that donors aim at increases in growth, a more careful allocation following the recommendations of our model should be able to improve outcomes with respect to growth. Future research might then want to investigate whether those parts of budget aid and project aid that are given in relation to informational advantages are indeed more effective in improving outcomes than those parts of such aid flows that are given due to other reasons. A differential analysis for (groups) of donors could also give additional insights as to which donors do and do not take account of information and bias, and whether these differences can explain potentially differential effects of these donors' aid. Finally, other types of delivering aid might also be investigated with respect to whether or not they are allocated in light of information and preferences. (38) We leave these questions for future research.

ABBREVIATIONS

CPIA: Country Policy and Institutional Assessment

CRS: Creditor Reporting System

DAC: Development Assistance Committee

ICRG: International Country Risk Guide

IMF: International Monetary Fund

OLS: Ordinary Least Squares

PPML: Poisson Pseudo Maximum Likelihood

UNGA: United Nations General Assembly

doi: 10.1111/ecin.12450

REFERENCES

Alesina, A., and D. Dollar. "Who Gives Foreign Aid to Whom and Why?" Journal of Economic Growth, 5, 2000, 33-63.

Alesina, A., and A. Drazen. "Why Are Stabilizations Delayed?" American Economic Review, 81, 1991, 1170-88.

Bailey, M. A., A. Strezhnev, and E. Voeten. "Estimating Dynamic State Preferences from United Nations Voting Data." Journal of Conflict Resolution, 61(2), 2017, 430-56.

Banks, A. S. Cross-National Time-Series Data Archive, 1815-2011. Jerusalem, Israel: Databanks International, 2011. Accessed March 15, 2017. http://www.cntsdata.com/

Bardhan, P. "Decentralization of Governance and Development." Journal of Economic Perspectives, 16, 2002, 185-205.

Barthel, F., E. Neumayer, P. Nunnenkamp, and P. Selaya. "Competition for Export Markets and the Allocation of Foreign Aid: The Role of Spatial Dependence among Donor Countries." World Development, 64, 2014, 350-65.

Basurto, P., P. Dupas, and J. Robinson. "Decentralization and Efficiency of Subsidy Targeting: Evidence from Chiefs in Rural Malawi." Working Paper, 2015.

Berthelemy, J.-C. "Bilateral Donors' Interest vs. Recipients' Development Motives in Aid Allocation: Do All Donors Behave the Same?" Review of Development Economics, 10, 2006, 179-94.

Besley, T., and T. Persson. Pillars of Prosperity: The Political Economics of Development Clusters. Princeton, NJ: Princeton University Press, 2011.

Boone, P. "Politics and the Effectiveness of Foreign Aid." European Economic Review, 40, 1996, 289-329.

von Borzyskowski, I. "Resisting Democracy Assistance: Who Seeks and Receives Technical Election Assistance?" Review of International Organizations, 11, 2016,247-82.

Bougheas, S., I. Dasgupta, and O. Morrissey. "Tough Love or Unconditional Charity?" Oxford Economic Papers, 59, 2007,561-82.

Chauvet, L., P. Collier, and A. Fuster. "Supervision and Project Performance: A Principal-Agent Approach." Working Paper, 2013.

Chung, K. C., P. Fleming, and E. Fleming. "The Impact of Information and Communication Technology on International Trade in Fruit and Vegetables in APEC." Asian-Pacific Economic Literature, 27(2), 2013, 117-30.

Claessens, S., D. Cassimon, and B. van Campenhout. "Evidence on Changes in Aid Allocation Criteria." World Bank Economic Review, 23(2), 2009, 185-208.

Clist, P., A. Isopi, and O. Morrissey. "Selectivity on Aid Modality: Determinants of Budget Support from Multilateral Donors." Review of International Organization, 7,2012, 267-84.

Cordelia, T., and G. dell'Ariccia. "Budget Support Versus Project Aid: A Theoretical Appraisal." Economic Journal, 117,2007, 1260-79.

Crawford, V., and J. Sobel. "Strategic Information Transmission." Econometrica, 50, 1982, 1431-51.

Dessein, W. "Authority and Communication in Organizations." Review of Economic Studies, 69, 2002, 811-38.

Dietrich, S. "Bypass or Engage? Explaining Donor Delivery Tactics in Foreign Aid Allocation." International Studies Quarterly, 57, 2013, 698-712.

Dixit, A. "Governance, Institutions and Economic Activity." American Economic Review, 99, 2009, 5-24.

Dollar, D., and V. Levin. "The Increasing Selectivity of Foreign Aid, 1984-2003." World Development, 34(12), 2006,2034-46.

Drazen, A. "Conditionality and Ownership in IMF Lending: A Political Economy Approach." IMF Staff Papers, 49, 2002,36-67.

Dreher, A. "Does Globalization Affect Growth? Evidence from a New Index of Globalization." Applied Economics, 38(10), 2006, 1091-110.

--. "IMF Conditionality: Theory and Evidence." Public Choice, 141(1-2), 2009, 233-67.

Dreher, A., and N. M. Jensen. "Country or Leader? Political Change and UN General Assembly Voting." European Journal of Political Economy, 29, 2013, 183-96.

Dreher, A., and S. Marchesi. "Information Transmission and Ownership Consolidation in Aid Programs." CESifo Working Paper No. 4437, 2013.

Dreher, A., J.-E. Sturm, and J. R. Vreeland. "Development Aid and International Politics: Does Membership on the UN Security Council Influence World Bank Decisions?" Journal of Development Economics, 88, 2009a, 1-18.

--. "Global Horse Trading: IMF Loans for Votes in the United Nations Security Council." European Economic Review, 53, 2009b, 742-57.

Dreher, A., P. Nunnenkamp, and R. Thiele. "Are 'New' Donors Different Comparing the Allocation of Bilateral Aid between NonDAC and DAC Donor Countries?" World Development, 39(11), 2011, 1950-68.

Dreher, A., A. Fuchs, R. Hodler, B. C. Parks, P. A. Raschky, and M. J. Tierney. "Aid on Demand: African Leaders and the Geography of China's Foreign Assistance." AidData Working Paper No. 3, 2014.

Dreher, A., K. Gehring, C. Kotsogiannis, and S. Marchesi. "Information Transmission within Federal Fiscal Architectures: Theory and Evidence." CEPR Discussion Paper No. 11344, 2016.

Dreher, A., V. Eichenauer, and K. Gehring. "Geopolitics, Aid and Growth: The Impact of UN Security Council Membership on the Effectiveness of Aid." World Bank Economic Review, Forthcoming.

Easterly, W. "Institutions: Top Down or Bottom Up?" American Economic Review: Papers and Proceedings, 98, 2008, 95-99.

Eichenauer, V. Z., and B. Reinsberg. "What Determines Earmarked Funding to International Development Organizations? Evidence from the New Multi-Bi Aid Data." Review of International Organizations, 12, 2017.

Fernandez, R., and D. Rodrik. "Resistance to Reform: Status Quo Bias in the Presence of Individual-Specific Uncertainty." American Economic Review, 81, 1991, 1146-55.

Fink, C., A. Mattoo, and I. C. Neagu. "Assessing the Impact of Communication Costs on International Trade." Journal of International Economics, 67, 2005, 428-45.

Fleck, R. K., and C. Kilby. "Changing Aid Regimes? US Foreign Aid from the Cold War to the War on Terror." Journal of Development Economics, 91, 2010, 185-97.

Franck, R., and I. Rainer. "Does the Leader's Ethnicity Matter? Ethnic Favoritism, Education, and Health in Sub-Saharan Africa." American Political Science Review, 106(2), 2012, 294-325.

Harris, M., and A. Raviv. "Allocation of Decision Making Authority." Review of Finance, 9, 2005, 353-83.

--. "A Theory of Board Control and Size." Review of Financial Studies, 21, 2008, 1797-832.

Hefeker, C. "Project Aid or Budget Aid? The Interest of Governments and Financial Institutions." Review of Development Economics, 10, 2006, 241-52.

Hoffler, A., and V. Outram. "Need, Merit, or Self-Interest: What Determines the Allocation of Aid?" Review of Development Economics, 15(2), 2011, 237-50.

Hollyer, J. R., B. P. Rosendorff, and J. R. Vreeland. "Democracy and Transparency." Journal of Politics, 73, 2011, 1194-205.

--. "Data Disclosure, Accountability and the Facets of Transparency." Working Paper, 2013.

--. "Measuring Transparency." Political Analysis, 22, 2014,413-34.

International Monetary Fund. Strengthening Country Ownership of Fund-Supported Programs. Washington, DC: International Monetary Fund, 2001.

Khan, M. S., and S. Sharma. "IMF Conditionality and Country Ownership of Programs." IMF Working Paper No. 142. Washington, DC: International Monetary Fund, 2001.

Koeberle, S. G. "Conditionality: Under What Conditions?" in Conditionality Revisited: Concepts, Experiences, and Lessons, edited by S. Koeberle, H. Bedoya, P. Silarszky, and G. Verheyen. Washington, DC: World Bank, 2005, 57-84.

Koeberle, S., Z. Stavreski, and J. Walliser, ed. Budget Support as More Effective Aid? Recent Experiences and Emerging. Washington, DC: World Bank, 2006.

Kuziemko, I., and E. Werker. "How Much Is a Seat on the Security Council Worth? Foreign Aid and Bribery at the United Nations." Journal of Political Economy, 114, 2006,905-30.

Lessmann, C., and G. Markwardt. "Decentralization and Foreign Aid Effectiveness: Do Aid Modality and Federal Design Matter in Poverty Alleviation?" CESifo Working Paper No. 3025, 2010.

Marchesi, S., L. Sabani, and A. Dreher. "Read My Lips: The Role of Information Transmission in Multilateral Reform Design." Journal of International Economics, 84,2011,86-98.

Milner, H. V., D. L. Nielson, and M. G. Findley. "Citizen Preferences and Public Goods: Comparing Preferences for Foreign Aid and Government Programs in Uganda." The Review of International Organizations, 11, 2016, 219-45.

Morrissey, O. "Fungibility, Prior Actions, and Eligibility for Budget Support," in Budget Support as More Effective Aid? Recent Experiences and Emerging Lessons, edited by S. Koeberle, Z. Stavreski, and J. Walliser. Washington, DC: World Bank, 2006. 333-47.

Mosley, R, and S. Abrar "Trust, Conditionality, and Aid Effectiveness," in Budget Support as More Effective Aid? Recent Experiences and Emerging Lessons, edited by S. Koeberle, Z. Stavreski, and J. Walliser. Washington, DC: World Bank, 2006, 311-32.

Nordtveit, I. "Herd Behavior or Increased Efficiency: What Determines the Allocation of General Budget Support?" Working Paper, University of Bergen, 2012.

OECD. "The Paris Declaration on Aid Effectiveness and the Accra Agenda for Action." 2005. Accessed March 15, 2017. www.oecd.org/dataoecd/11/41/34428351.pdf.

--. "Creditor Reporting System." 2016. Accessed June 15, 2016. http://stats.oecd.org/Index.aspx?datasetcode=CRS 1.

Ouattara, B., and E. Strobl. "Aid, Policy and Growth: Does Aid Modality Matter?" Review of World Economics, 144, 2008, 347-65.

PRS Group. International Country Risk Guide (ICRG). Syracuse, NY: IBC, 1998, http://www.prsgroup.com.

Rajan, R., and A. Subramanian. "Aid and Growth: What Does the Evidence Really Show?" Review of Economics and Statistics, 90, 2008, 643-65.

Reinsberg, B. "Organizational Reform and the Rise of Trust Funds: Lessons from the World Bank." Review of International Organizations, 12, 2017.

Rommel, T., and P. Schaudt. "First Impressions: How Leader Changes Affect Bilateral Aid." CESifo Working Paper No. 6047, 2016.

Sigelman, L., and L. Zeng. "Analyzing Censored and Sample-Selected Data with Tobit and Heckit Models." Political Analysis, 8(2), 1999, 167-82.

Strezhnev, A., and E. Voeten. "United Nations General Assembly Voting Data." 2013. Accessed March 15, 2017. http://hdl.handle.net/1902.1/12379 UNF:5:NpHV5DXWPNWMWOrLGTjQYA== Erik Voeten [Distributor] V5 [Version].

Tabellini, G., and A. Alesina. "Voting on the Budget Deficit." American Economic Review, 80, 1990, 37-49.

Van de Sijpe, N. "Is Foreign Aid Fungible? Evidence from the Education and Health Sectors." World Bank Economic Review, 27,2013, 320-56.

Voeten, E. "Clashes in the Assembly." International Organization, 54, 2000, 185-215.

Williams, A. "A Global Index of Information Transparency and Accountability." Journal of Comparative Economics, 43(3), 2015, 804-24.

Winters, M., and G. Martinez. "The Role of Governance in Determining Foreign Aid Flow Composition." World Development, 66, 2015, 516-31.

World Bank. "World Development Indicators." CD-Rom, Washington, DC, 2013. Accessed September 4, 2013. http://data.worldbank.org/data-catalog/world-d.

SUPPORTING INFORMATION

Additional Supporting Information may be found in the online version of this article:

Appendix S1. Definition and properties of the communication game

Appendix S2. Derivation of donor and recipient government's ex ante expected losses

Appendix S3. Proof of statement in Section III

Appendix S4. Sources and definitions

Appendix S5. Descriptive statistics (estimation sample, column 5 (6) in Table 1)

Appendix S6. Correlations

Appendix S7. Additional figures

AXEL DREHER, SARAH LANGLOTZ and SILVIA MARCHESI *

* We thank Christian Bj0rnskov, Oliver Morrissey, Ola Olsson, Andrea Presbitero, Bernhard Reinsberg, Jonathan Temple, participants at the Beyond Basic Questions Workshop (Salzburg 2016), and the Centre for the Study of African Economics Conference (CSAE, Oxford 2014) for helpful comments, and Jamie Parsons for proof-reading.

Dreher: Professor, Alfred-Weber-Institute for Economics, Heidelberg University, Heidelberg 69115, Germany. Phone +49 6221 54 2921, Fax +49 6221 54 3649, E-mail mail@axel-dreher.de

Langlotz: Dipl.-Vw., Alfred-Weber-Institute for Economics, Heidelberg University, Heidelberg 69115, Germany. Phone +49 6221 54 3172, Fax +49 6221 54 3649, E-mail sarah.langlotz@awi.uni-heideIberg.de

Marchesi: Professor, Dipartimento di Economia, Metodi Quantitativi e Strategic di Impresa (DEMS), Universita di Milano Bicocca, Milano 20126, Italy. Phone +39 02 6448 3057, Fax +39 02 6448 3085, E-mail silvia.marchesi@unimib.it

(1.) In the words of Koeberle (2005, 67) ownership is the "commitment to aid-supported reforms by country authorities and a majority of domestic stakeholders." According to Khan and Sharma (2001,13) ownership "refers to a situation in which the policy content of the program is similar to what the country itself would have chosen." The International Monetary Fund (IMF 2001,6) defines it as "a willing assumption of responsibility for an agreed program of policies, by officials in a borrowing country who have the responsibility to formulate and carry out those policies, based on an understanding that the program is achievable and is in the country's own interest."

(2.) The mechanisms and circumstances under which such information should be transferred by recipient countries to donors (or lenders) have rarely been investigated. An exception is Marchesi, Sabani, and Dreher (2011) who analyzed the specific case of communication between the IMF and a borrowing country.

(3.) Koeberle, Stavreski, and Walliser (2006) emphasize that budget support comes with greater country ownership and higher spending on services that countries prioritize in their own budgets. This does not imply that the aid transfer is necessarily unconditional. However, bilateral donors do not usually condition their aid on a large number of detailed conditions. If conditions are attached to the aid they usually refer to the general stance of a country's policy, e.g., with respect to human rights conditions, democracy, or the absence of corruption.

(4.) In a dynamic framework, Bougheas, Dasgupta, and Morrissey (2007) also address the choice between conditional and unconditional transfers. They show that conditionality is self-perpetuating even when it is not optimal. The results in Bougheas et al. thus question the wisdom of conditionality at large. Also see Dreher (2009).

(5.) More generally, the relationship between decentralization and development has been analyzed, among others, by Bardhan (2002) and Lessmann and Markwardt (2010).

(6.) More recently, Basurto, Dupas, and Robinson (2015) have shown that a decentralized allocation of subsidies in rural Malawi may offer informational advantages, despite of being prone to elite capture.

(7.) An exception is Marchesi, Sabani, and Dreher (2011), who--building on the cheap talk literature (Crawford and Sobel 1982; Dessein 2002; Harris and Raviv 2005, 2008)--have identified and tested the conditions under which it is optimal for the IMF to delegate control to a recipient country in order to maximize the quality of a reform program. More recently, and in a different context, Dreher et al. (2016) explore the role of information transmission in explaining the optimal degree of decentralization across countries.

(8.) For exceptions see Berthelemy (2006), Claessens, Cassimon, and van Campenhout (2009), Dreher, Nunnenkamp, and Thiele (2011), Dietrich (2013), Barthel et al. (2014), and Winters and Martinez (2015).

(9.) Specifically, Appendix S1 defines and shows the properties of the communication game, Appendix S2 derives the ex ante expected losses of the federal and local governments, while Appendix S3 contains proofs of the statements made in Section III.A.

(10.) The analytics feature the case in which the donor cannot commit to an incentive-compatible decision rule in which the revelation principle applies. This assumption fits in well with the specific relationship between a donor and a recipient government in which the principal cannot use a standard mechanism to elicit private information from the agent.

(11.) The utility function (Equation (1)) can be derived from a more general objective function [[??].sup.D] = Y(p) + [gamma]E(p), where Y is the recipient country's output and E(p) measures the externalities that the choice of a specific policy p could have on the donor country's economy. The parameter [gamma] (0 [less than or equal to] [gamma] [less than or equal to] 1) denotes the importance of such spillover effects. Specifically, if the recipient country is big, [gamma] will tend to 1, while for very small countries [gamma] will be close to 0. Taking a Taylor expansion of [[??].sup.D] (p) up to the second term, one obtains the form in Equation (1).

(12.) The more general function is [[??].sup.G] = Y(p) + [theta]C(p), where C are contributions from special interest groups. We assume that C decreases with p and that the parameter [theta] (O [less than or equal to]theta] [less than or equal to] 1) denotes the importance of lobbies. Using a Taylor expansion of [[??].sup.G](p) up to the second term, one obtains Equation (2).

(13.) We should emphasize that our scenario of conflicting preferences does not characterize all donor-recipient relationships, as there could be cases where preferences over core policies are relatively closely aligned. Even in such cases, however, contentious issues are likely and each party owns some private information so that the essence of the model keeps to be applicable.

(14.) This is to some extent similar to how Bougheas, Dasgupta, and Morrissey (2007) relate the choice of conditionality to donors' priors on recipient types.

(15.) Since the derivative of D(G,B) with respect to B cannot be analytically derived, this result is obtained by numerical simulations (see Harris and Raviv 2005).

(16.) Some donors make their aid conditional on the implementation of certain policies, or on the absence of corruption, human rights violations, or restrictions on democracy. This holds in particular for multilateral donors like the IMF or the World Bank that we do not focus on in this paper. Some bilateral aid agencies also attach conditions. For example, the United States' Millennium Challenge Corporation conditions its aid on the absence of corruption, government effectiveness, and low inflation, among others. This does not restrict the recipient governments on deciding about what to use the aid for, however. The same holds for those parts of aid that are restricted to be spent in the donor country (so-called tied aid). While tying aid reduces its value for the recipient, the recipient government is free to decide about which projects or purposes to use the aid for.

(17.) This assumes that aid is not fully fungible, which is supported by the recent literature. For example, Van de Sijpe (2013, 26) shows "little evidence that aid is fully or even largely fungible; rather, most point estimates suggest limited fungibility." See also Milner, Nielson, and Findley (2016). As we focus on donor choices rather than recipient country policies it is sufficient for our analysis that donors expect aid not to be fully fungible, independent of whether or not it is in fact fungible after the donors delivered the aid.

(18.) For example, the relative share of "soft" to "hard" information is likely to depend on the quality of the communication infrastructure.

(19.) This is particularly severe for budget aid, with less than 900 non-zero dyad-years in our sample (see Appendix S5 for details and Winters and Martinez 2015 for a similar approach). The inclusion of dyad-specific fixed effects ensures that donor-recipient pairs with no positive aid over the entire sample period do not contribute to the estimates.

(20.) Note that the share of project and budget aid to overall commitments is substantially larger than one for some observations. The reason is that overall commitment data provided by the OECD include repayments, while project and budget aid commitments do not include repayments. When we replicate our main regressions excluding those observations our results are broadly unchanged. As a second test for robustness, we compute overall gross commitments based on data for the subcategories. Results again remain robust. Also note that while the mean values for project aid and budget aid are rather low in our sample due to a large number of zero donor-recipient-year observations, project aid is on average 70% of total commitments when we focus on nonzero amounts, while the share of budget aid is close to 30% (see Appendix S5 for summary statistics including and excluding zero-observations). The two do not sum up to total commitments, which also include donors' administrative costs, unallocated or unspecified aid, aid for refugees in donor countries, developmental food aid, other commodity assistance, debt relief, and humanitarian aid.

(21.) In Tabellini and Alesina (1990), given a situation of political instability and polarization, a balanced budget does not represent a political equilibrium. This is because the current majority does not internalize the costs of budget deficits and the more this is the case, the greater the difference between its preferences and the expected preferences of future majorities. Alesina and Drazen (1991) find that when stabilization has significant distributional implications a "war of attrition" among different socioeconomic groups may delay stabilization.

(22.) We alternatively include a binary indicator for Autonomous Regions and the share of Subnational Expen ditures/Total Expenditures as proxies for stronger regional vetoes and thus a larger status quo bias.

(23.) We should emphasize that Ethnic Tensions could at the same time be related to the importance of the recipient's private information (when a country is more multifaceted from a social point of view its local knowledge is more important).

(24.) As an alternative measure for the dyadic bias we use a binary variable on Democratic Distance relying on the Polity IV index of democracy. We calculate Democratic Distance between the donor and recipient as one if either the donor or the recipient is a democracy (i.e., a Polity IV index larger than five), while the respective other country is not a democracy (i.e., has a value below six). We expect the agency bias to be smaller among democracies, as democratic countries tend to agree on a broad set of principles regarding political and economic liberalism (Voeten 2000). Furthermore, we use an indicator of Ideological Distance, measured as the absolute difference between the donor and recipient government on a left-to-right spectrum.

(25.) Missing data entries can result from a number of reasons. For example, (1) the recipient government might have the data but does not report them, (2) the recipient government does not have the information, or (3) the recipient reports the data to the World Bank but Bank staff choses to not report them, for example, because they consider them insufficiently reliable. In all these cases, missing data proxy for intrans-parent environments that make the recipient's private information more important relative to the donor's. This would not be the case if data that have been missing at the time the decisions about how to give aid have been made had later been included to the database. In this case we would report Transparency to be too high. However, the congelation between our missing indicator variable and an indicator constructed in analogy based on an earlier--2005--version of the World Development Indicators is very high. Correlation between our indicator and those of Hollyer, Rosendorff, and Vreeland (2011) is 0.80 and our results are robust to using their index instead of ours. Our indicator is also significantly correlated with the HRV Index of transparency (Hollyer, Rosendorff, and Vreeland 2014), which uses patterns in the missing data to model transparency as a latent variable, and a Combined Transparency indicator based on 29 sources taken from Williams (2015).

(26.) Media Access is a composite indicator including access to TV, radio, papers, and internet (using data from Banks 2011). Internet Users and Telephone Lines are also highly correlated (rho=0.64), but sample size is reduced substantially when we include Internet Users. We test robustness to using Newspapers in circulation (per 1,000 inhabitants) and the number of Internet Users per 100 people. Interestingly, the correlation between Telephone Lines and Transparency is weak, indicating that these measures account for different aspects of transparency (see Hollyer, Rosendorff, and Vreeland 2013 for a detailed discussion of these differences). We therefore include these two measures at the same time rather than separately.

(27.) Due to data availability, we compute the number from 1970, leading to a maximum experience of 40 years.

(28.) We tested robustness by including other control variables. We included the World Bank's Country Policy and Institutional Assessment (CPIA) in order to control for "recipient merit." We included the recipient country's KOF Index of Globalization (Dreher 2006) to capture its general openness. We controlled for those types of aid that are neither project aid nor budget aid (e.g., humanitarian aid or aid related to refugees). What is more, we also tested robustness to excluding GDP and population. None of our results is affected by these changes.

(29.) In a previous version of this article, we estimated our regressions using Poisson pseudo maximum likelihood (PPML) at the recipient-country-level, with project aid and budget aid in levels rather than shares (Dreher and Marchesi 2013). Results were in line with the model's predictions. When we run PPML in our dyadic fixed effects model, the incidental parameter problem becomes paramount, and models do not converge. One might also think of using models such as Tobit or Heckman--two commonly used methods in the aid allocation literature--but their use would be problematic with our data (see Sigelman and Zeng 1999). Tobit may lead to biased estimates when zero observations are not the result of censoring mechanisms, while Heckman is inefficient when the dependent variable is exclusively nonnegative. What is more, in our short sample the dyadic-fixed effects Tobit estimates are biased due to the incidental parameter problem. Crucially, the Tobit and Heckman models do not converge with our dyadic-fixed effects. When we instead run seperate models for the selection of aid recipients and the allocation of aid amounts (including dyadic fixed effects), we find strong results in line with the model at the selection stage, but weaker results when it comes to allocation. When we estimate a Heckman Selection model with donor-, year-, and recipient-fixed effects instead of dyadic fixed effects results are similar. Information and preference divergence thus seem to be more important when deciding about whether to give aid at all, rather than deciding about the amount of aid. When we run Tobit with donor-, year-, and recipientfixed effects results are overall in line with the model's predictions. In order to better capture long-term relationships between donors and recipients, we have also run specifications (1) excluding country-pair-fixed effects, (2) including fixed-effects for recipients rather than country-pairs, and (3) including donor- and recipient-fixed effects in tandem. Our overall results are robust to these changes. We also included recipient-year- and donor-year-fixed effects in addition to the country-pair-fixed effects. Unsurprisingly, most coefficients are no longer significant at conventional levels in this specification, with the exception of those variables that vary at the recipient-donor-year-level. Results are robust when we include recipient-year- and donor-year-fixed effects instead of the country-pair-fixed effects.

(30.) See, for example, Alesina and Dollar (2000), Dollar and Levin (2006), and Nordtveit (2012).

(31.) Again, Alesina and Dollar (2000), or Dollar and Levin (2006) are useful examples. Also see Dreher, Nun nenkamp, and Thiele (2011).

(32.) In column 5, while the coefficient of transparency becomes marginally insignificant, its magnitude remains much larger for project than budget aid, consistent with the results of columns 1 to 4. Similarly, in column 3 of Table 2, the coefficient of transparency becomes insignificant but is still much larger for project aid than budget aid.

(33.) In a robustness test, we exclude repayments from total commitments, such that the shares do not exceed 100%. Transparency then remains statistically significant at the 10% level in columns 3 and 5. The effect also remains statistically significant in columns 3 and 5 (at the 1% level) when we replace the dependent variable with the log of project and budget aid in constant USD (rather than using shares).

(34.) We considered the alternative monadic proxies Autonomous Regions and Subnational Expenditures/Total Expenditures and the dyadic proxies Democratic Distance and Ideological Distance and also obtained no robust results for these bias-related variables in either direction, on average.

(35.) We should stress here that the influence of the agency bias on the amount of budget aid and project aid is also consistent with the (theroretical) results of Cordelia and dell'Ariccia (2007) who find that budget support should be preferred to project aid when the donor's preferences are close to those of the recipient.

(36.) Appendix S7 shows similar figures focusing on the interaction between UNGA Distance and Telephone Lines. Overall, results are similar to those shown in Figures 2 and 3.

(37.) As shown by Dreher, Eichenauer, and Gehring (Forthcoming), donors' geopolitical motives for grantig aid reduce the effectiveness of aid in increasing economic growth.

(38.) As one example, our model could be used to explain the increasing amount of aid that is channeled via multilateral institutions as non-core aid ("multi-bi aid"), see Eichenauer and Reinsberg (2017) and Reinsberg (2017).

Caption: FIGURE 1 Choice among Centralization and Delegation as a Function of D and G

Caption: FIGURE 2 Marginal Effect of UNGA Distance as Transparency Changes

Caption: FIGURE 3 Marginal Effect of UNGA Distance as Transparency Changes

TABLE 1

Main Results

                         (1)          (2)           (3)
                     Project Aid   Budget Aid   Project Aid

(log) GDP per          -0.0407     0.0062 ***     -0.0412
capita (t - 1)        (0.2900)      (0.0097)     (0.3037)

(log) Population       0.0108      0.0075 **     -0.0798 *
(t - 1)               (0.8090)      (0.0464)     (0.0806)

Transparency         0.3538 ***    0.0178 **     0.2498 *
(t - 1)               (0.0069)      (0.0211)     (0.0211)

Telephone            0.0073 ***     -0.0001     0.0058 ***
lines (t- 1)          (0.0000)      (0.2497)     (0.0000)

Donor experience     0.0019 ***     -0.0000     0.0019 ***
(t - 1)               (0.0000)      (0.8831)     (0.0000)

Bureaucratic           -0.0013      -0.0001       -0.0057
quality (t - 1)       (0.8997)      (0.9005)     (0.5831)

Bilateral                                       0.0150 ***
history (t - 1)                                  (0.0000)

(log) Bilateral                                   -0.0007
trade (f - 1)                                    (0.5567)

Government             0.0032        0.0003       0.0031
capability            (0.5359)      (0.3357)     (0.5549)
(t - 1)

Ethnic tensions        0.0032       -0.0002       0.0022
(t- 1)                (0.7401)      (0.6848)     (0.8224)

UNGA distance
(t - 1)

Adj. R-squared          6.004        0.001         0.004

F Statistic            32.658        2.005        32.783

Number of               46378        46378         46014
observations

Number of               3126          3126         3126
country pairs

Number of                112          112           112
recipients

                         (4)           (5)           (6)
                     Budget Aid    Project Aid   Budget Aid

(log) GDP per        0.0069 ***      -0.0280     0.0088 ***
capita (t - 1)        (0.0066)      (0.5111)      (0.0032)

(log) Population      0.0070 *      -0.0885 *     0.0078**
(t - 1)               (0.0594)      (0.0576)      (0.0388)

Transparency          0.0167 **      0.1913       0.0136 *
(t - 1)               (0.0274)      (0.1237)      (0.0880)

Telephone              -0.0001     0.0052 ***      -0.0001
lines (t- 1)          (0.1486)      (0.0001)      (0.1025)

Donor experience       -0.0000     0.0020 ***      -0.0000
(t - 1)               (0.8440)      (0.0000)      (0.9207)

Bureaucratic           -0.0001       -0.0035       -0.0003
quality (t - 1)       (0.8448)      (0.7346)      (0.6522)

Bilateral             0.0002 *     0.0153 ***      0.0001
history (t - 1)       (0.0933)      (0.0000)      (0.1901)

(log) Bilateral      -0.0002 ***     -0.0007     -0.0003 ***
trade (f - 1)         (0.0078)      (0.5197)      (0.0060)

Government             0.0003        0.0024        0.0003
capability            (0.3504)      (0.6593)      (0.2516)
(t - 1)

Ethnic tensions        -0.0002       0.0061        -0.0002
(t- 1)                (0.7595)      (0.5182))     (0.7223)

UNGA distance                        -0.0012      -0.0024 *
(t - 1)                             (0.9748)      (0.0588)

Adj. R-squared          0.001         0.005         0.001

F Statistic             2.074        32.261         2.249

Number of               46014         43944         43944
observations

Number of               3126          3039          3039
country pairs

Number of                112           109           109
recipients

Notes: OLS at the donor-recipient-year level. Donor-recipient
-fixed-and year-fixed effects are included. Standard errors
are in parentheses (clustered at the donor-recipient level;
significance levels; * 0.10, ** 0.05, *** 0.01).

TABLE 2

Interaction Effects

                                     (1)            (2)
                                 Project Aid     Budget Aid

(log) GDP per capita (t - 1)       -0.0300       0.0090 ***
                                  (0.4769)        (0.0026)

(log) Population (t - 1)          -0.0755 *        0.0061
                                  (0.0914)        (0.1003)

Transparency (t - 1)              0.5143 **    -0.0300.0299 *
                                  (0.0405)        (0.0711)

Telephone lines (t - 1)          0.0050 ***       -0.0001
                                  (0.0004)        (0.1990)

Donor experience (t - 1)         0.0020 ***        0.0000
                                  (0.0000)        (0.9326)

Bureaucratic quality (t - 1)       -0.0032        -0.0003
                                  (0.7560)        (0.6112)

Bilateral history (t - 1)        0.0157 ***        0.0001
                                  (0.0000)        (0.4343)

(log) Bilateral trade (t - 1)      -0.0007      -0.0003 ***
                                  (0.5486)        (0.0052)

Government capability (t - 1)      0.0029          0.0002
                                  (0.5947)        (0.3738)

Ethnic tensions (t - 1)            0.0062         -0.0002
                                  (0.5096)        (0.6937)

UNGA distance (t - 1)             0.1244 **     -0.0193 ***
                                  (0.0141)        (0.0028)

UNGA*Transparency (t - 1)         -0.1971 *      0.0265 ***
                                  (0.0725)        (0.0082)

UNGA*Telephone lines (t - 1)

Adj. R-squared                      0.005          0.001

F Statistic                        31.523          2.170

Number of observations              43944          43944

Number of country pairs             3039            3039

Number of recipients                 109            109

                                     (3)           (4)
                                 Project Aid   Budget Aid

(log) GDP per capita (t - 1)       -0.0269     0.0083 ***
                                  (0.5137)      (0.0050)

(log) Population (t - 1)          -0.0875 *     0.0074 **
                                  (0.0582)      (0.0493)

Transparency (t - 1)               0.1869       0.0154 *
                                  (0.1399)      (0.0546)

Telephone lines (t - 1)           0.0057 **    -0.0003 ***
                                  (0.0127)      (0.0022)

Donor experience (t - 1)         0.0020 ***      -0.0000
                                  (0.0000)      (0.9822)

Bureaucratic quality (t - 1)       -0.0036       -0.0003
                                  (0.7266)      (0.6991)

Bilateral history (t - 1)        0.0153 ***      0.0001
                                  (0.0000)       0.0007

(log) Bilateral trade (t - 1)      -0.0007     -0.0003 ***
                                  (0.5169)      (0.0068)

Government capability (t - 1)      0.0024        0.0003
                                  (0.6545)      (0.2751)

Ethnic tensions (t - 1)            0.0062        -0.0002
                                  (0.5035)      (0.6622)

UNGA distance (t - 1)              0.0018      -0.0036 **
                                  (0.9657)      (0.0153)

UNGA*Transparency (t - 1)

UNGA*Telephone lines (t - 1)       -0.0003     0.0001 ***
                                  (0.7552)      (0.0065)

Adj. R-squared                      0.005         0.001

F Statistic                        31.231         2.171

Number of observations              43944         43944

Number of country pairs             3039          3039

Number of recipients                 109           109

Notes: OLS at the donor-recipient-year level. Donor-recipient
-fixed-and year-fixed effects are included. Standard errors are
in parentheses (clustered at the donor-recipient-level;
significance levels: * 0.10, ** 0.05, *** 0.01).