Katangale or kuba?: development assistance and the experiences and perceptions of local corruption in Malawi

ABSTRACT

There is lively debate concerning the influence of development assistance (‘aid’) on corruption in recipient countries; however, to date, few studies have systematically examined the subject sub-nationally. This study estimates the association between sub-national aid levels and citizens’ perceptions and experiences of local corruption in Malawi. Overall, we find that individuals in districts receiving high amounts of aid are no more likely to view local leaders as corrupt than those in districts receiving lower amounts of aid. However, higher levels of aid are associated with more experiences of bribe solicitation. We also find evidence that aid channeled through NGOs may have different effects than government-implemented aid, as it is associated with better perceptions of local corruption and reduced bribe solicitation. The findings suggest that, in aggregate, corruption resulting from receiving aid may not be of sufficient magnitude or visibility to change citizens’ broader beliefs about government performance or legitimacy.

KEYWORDS:

• Malawi
• corruption
• development assistance
• aid

Introduction and review

This study examines the relationship between levels of development assistance (hereafter ‘aid’) and citizens’ perceptions and experiences of local corruption in Malawi. Corruption is an overriding concern among those who study, fund, and implement aid projects.¹ Indeed, it is rare to find a study on the economic or political impact of aid that does not mention corruption. The study of corruption in aid is motivated not only by a normative, moral objection to the abuse of entrusted power in aid projects, but also by the belief that corrupt practices erode the effectiveness of aid by diverting resources away from where they could have the most impact. In addition, there is a well-founded concern that corruption in public services erodes government legitimacy and suppresses political participation in heavily aid-dependent states.

Aid resources are thought to be particularly prone to corruption because the allocation and implementation of aid projects within receiving countries is largely left to the discretion of local leaders; because these local leaders often exercise monopolies over valuable aid resources; and because systems of accountability for aid programs are often weak (Klitgaard 1988). Local leaders monopolize the management of aid resources at the community level. The combination of ambiguous and overlapping organizational jurisdictions, a multiplicity of project implementers, and the great distance of donors from aid projects impedes accountability. Donors, for their part, have little incentive or capacity to actively monitor projects or to punish misappropriation, largely because of cultural ties, a desire to maintain trading relationships, and foreign policy interests at home (Alesina and Dollar 1998; Tavares 2003; Williamson 2010).²

Although most studies of the relationship between aid and corruption focus on national actors, it is reasonable to assume that the potential for corruption in aid projects is also significant at the local, community level, where discretion is greatest. At the community level, aid projects are usually carried out by local government officials, national NGOs (non-governmental organizations), or community- and faith-based organizations (CBOs), such as mission hospitals and self-help groups. These intermediary organizations and the local leaders that manage them are a crucial part of the ‘aid chain.’ They direct and manage a great deal of the local distribution of aid resources (Bierschenk, Elwert, and Kohnert 1993; Neubert 1996; Dionne 2012)

The centrality of these local leaders in aid projects rests on the fact that donors, international NGOs, and, to a lesser extent, national governments, usually lack the cultural understanding, knowledge of the local political context, language skills, or the physical presence necessary to carry out projects at the community level (Lewis and Mosse 2006). As a result, national and international agencies depend heavily on intermediaries to ‘navigate local barriers’ for them (Swidler 2009). For community members, local leaders such as traditional leaders and the head of CBOs therefore, act as ‘gatekeepers, conduits of information, and respected arbiters’, and are the face of aid activity in the community (Dionne 2012). For example, they decide who will serve on project committees and which community members are considered to be ‘vulnerable’ and therefore should be targeted for support. They may also decide where infrastructure such as water pumps or schools should be placed. Local leaders also organize the communal and ‘voluntary’ activity often required to implement these projects’ (Swidler 2010). Any blame or credit for the way in which aid projects are managed should, therefore, accrue to them disproportionately.

Aid and perceptions of corruption: the evidence

Over sixty years of experience with aid projects has given us only tenuous empirical evidence that influxes of aid resources are associated with increased corruption in receiving communities. Anecdotal evidence of the mismanagement and misuse of development aid resources abounds (Moyo 2009; Rayner and Swinford 2011). Cross-national studies also provide some evidence of a direct, positive association between aid flows and national levels of corruption (Svensson 2000; Knack 2001), although this evidence is mixed (Alesina and Weder 2002; Tavares 2003; Okada and Samreth 2012). There is also cross-national evidence that aid infusions can increase resources available for patronage, skew incentives within social service bureaucracies, and promote rent seeking (Brautigam 2000; Van de Walle 2001; Brautigam and Knack 2004; Harford and Klein 2005; Moss, Pettersson, and Van de Walle 2006; Djankov, Montalvo, and Reynal-Querol 2008).

Although the scope for discretion, and therefore, corruption, is large at the community level, there are few rigorous sub-national studies of the relationship between aid levels and the prevalence of community-level corruption. We have limited knowledge about the extent of corruption that results from aid projects at the community level or how such projects are perceived by local beneficiaries in regard to corruption. On the few occasions when project beneficiaries have been surveyed in a systematic manner regarding the management of aid projects, they were found to have perceived high levels of corruption in aid projects (Bailey 2008; Olken 2009; Anderson, Brown, and Jean 2012). In one such study, community members were reported to be extremely cognizant of waste in aid projects, noting, for example, that project money spent on workshops, travel, and training and the ‘high salaries, expensive housing and cars, fringe benefits, and ostentatious consumption’ could have been better spent on the direct provision of goods and services (Anderson, Brown, and Jean 2012, 107). Overall, community members have been found to be knowledgeable about the ways in which projects can be manipulated by local leaders for personal gain and they are cynical about the integrity of the CBOs that implement them.

The relationship between communities’ perceptions of corruption and the actual prevalence of resource misappropriation and corruption in aid projects is unclear. In the qualitative studies discussed above, communities’ perceptions of corruption were not necessarily based on actual experiences of corruption. Rather they seemed to emerge from vague rumors based on ‘suspicions’ and a feeling that ‘aid has helped too few people or accomplished too little given the quantity of resources they know has been provided’ (Anderson, Brown, and Jean 2012). These community-level findings were in keeping with empirical cross-national studies of corruption that found only tenuous associations between national-level citizen perceptions of corruption and the actual prevalence of corrupt practices (Seligson 2002; Treisman 2007).³ To summarize, in-depth qualitative studies find that regardless of the actual incidence of corruption in aid projects, aid activity can generate a diffuse suspicion and a feeling that aid resources are being misappropriated.

Case studies and evaluations of participatory, community-based aid projects have similar findings. These studies find that the presence of aid projects in a community is sometimes associated with increased community tensions and conflicts over resources. For example, in her detailed case study of HIV/AIDS care and support organizations in peri-urban Tanzania, Jelke Boesten (2011) found that the infusion of HIV/AIDS funding for community-based programs into a setting of scarce resources and insecurity increased community tensions and led to allegations of corruption. Several rigorous, large-scale project evaluations of community-driven development projects have also found evidence that aid projects can increase community discord and encourage the community members to view the local leaders as corrupt (Chase and Woolcock 2005; Barron, Woolcock, and Diprose 2007; King, Samii, and Snilstveit 2010). While the findings from these studies may be accurate in particular sectors and in particular points in time, their empirical base rests almost exclusively on case studies and evaluations of individual projects. It is therefore unclear whether the perceptions that these individual aid projects generate will still persist when we look at aid activity in aggregate, over time.

Patronage and local understandings of corruption in Malawi

The association between the misuse of aid resources and the perceptions of corruption in local leaders may not be straightforward because corrupt practices intersect with, and occur through, culturally legitimate, pre-existing social practices. In the sub-Saharan African context, some forms of corruption are seen as a consequence of clientelistic political systems based on patron-client networks and ‘informal institutions of reciprocity’ (MacLean 2011; Smith 2014). In these systems, a client exchanges loyalty, political support, and service for personal or communal material benefits such as jobs or food supplies from a patron.

Such patron-client interactions are understood to be part of long-term relationships that form ongoing and far-reaching networks of obligations, rather than being simple one-off instrumental exchanges. These relationships are deeply embedded in the social systems of family, clan, and religious community and are bounded by profound moral imperatives (Oliver de Sardin 1999; Blundo and de Sardan 2006; Harrison 2006; Swidler 2009). They reflect cultural patterns rooted in the geography and economics of African countries where ‘wealth in people – children, wives, clients, and other dependents – [is] the key to increasing material wealth’ i.e. where having a network of clients has been and continues to be important for survival (Kopytoff and Miers 1977 cited in Swidler 2009). The inability to participate in a patron-client network, e.g. not being able to help kin members in need, can cause deep moral crisis, and can lead to stigmatization and retribution as well as a loss of political and social status (Chabal and Daloz 1999; Anders 2002; Harrison 2006; MacLean 2011; Smith 2014).⁴

In these settings, local practices that may look like an abuse of power from the outside, such as nepotism and ethnic favoritism may be construed as upholding ones’ obligations and behaving responsibly. Local perceptions of corruption may be flexible and context-specific, depending on the relationship between the actors involved (Sissener 2001; Blundo and de Sardan 2006; Harrison 2006). What may drive perceptions of corruption in this context is not that resources are used for unintended purposes, or that they are distributed through informal networks, but rather whether or not they are shared widely and equitably within these networks (Chabal and Daloz 1999; Sissener 2001; Blundo and de Sardan 2006; Smith 2007; Smith 2014).

These varying moral understandings of corrupt practices are evident when we examine beliefs about corruption in Malawi specifically. Historically, Malawi has been considered to be one of the least corrupt countries in sub-Saharan Africa. This changed after Hastings Banda, the former President for Life, was ousted in a democratic transition in the mid-1990s. Both information about corruption and the practice of corruption itself grew due to the absence of top-down control, weak oversight power within the newly formed political parties, and the perceived need within the new leadership to build political coalitions through the distribution of resources (Anders 2002; Cammack, Kelsall, and Booth 2010). Banda's regime may not have been less corrupt than his successors’, but the type and visibility of corruption differed under his administration (Hall-Matthews 2007). Under President Muluzi, corruption scandals, often involving government procurement, became a recurring feature of political life as did anti-corruption drives, heavily sponsored by external funders (Booth et al. 2006). These anti-corruption drives intensified under President Mutharika when they were often used instrumentally to delegitimize and hamstring political opponents (Hall-Matthews 2007).

Citizens in Malawi perceive high levels of corruption across government agencies and view such practices negatively (Khaila and Chibwana 2005; Chinsinga, Kayuni, and Konyani 2010). Frustration with corruption, particularly with large-scale procurement scams is a regular topic of newspaper articles, call-in radio shows, popular song (Hussein 2005; Chirambo 2009; Phiri 2016); and tackling corruption is a central demand in anti-government protest (Cammack 2012). There is also convincing evidence that disgust with President Joyce Banda's handling of the massive ‘Cashgate’ scandal in which MK 20 billion was skimmed from central government accounts played a significant role in her loss in Malawi’s 2014 elections (Zimmerman 2015).

Despite this widespread public disapproval, anthropologists studying corruption in Malawi and elsewhere on the continent have consistently found that attitudes about benefiting from corruption (particularly lower-level, petty corruption) are often nuanced and ambivalent. For example, in his anthropological study of corruption in Malawi's civil service, Gehard Anders (2002) noted that when one examined the ‘everyday language’ used to discuss corruption in Malawi there is no ‘clear and unequivocal rejection of practices considered to be corrupt among average Malawians’ (12). He found that citizens use a variety of terms to describe illegal, and informal work practices. One, ‘katangale’, covers ‘dubious or shady’ practices linked to work including patronage and nepotism. He states that this term has strong, quasi-spiritual links to the idea of obligatory sharing and reciprocity and that benefiting from katangale is not always considered bad; rather it might be appropriate and in keeping with the moral order. This term can be distinguished from ‘kuba’, theft, which is used to describe the use of resources for one's personal benefit with no implications of proper distribution. In all situations, kuba is considered morally wrong.⁵ This nuanced attitude may be particularly prevalent in times of economic upheaval and in situations of pervasive corruption when official avenues to reliably access public goods are scarce and/or closed.

Daniel Jordan Smith's (2014) ethnographic study of HIV/AIDS in Nigeria found similar ambivalence in staff attitudes towards misappropriation of funds by NGO directors. He found that when directors acted as ‘good patrons’ who were flexible and generous to staff in decisions about the personal use of NGO funds, they were largely excused; in contrast, directors perceived as using NGO funds for narrow personal gain and self-aggrandizement were ostracized (Smith 2014).⁶

In addition to being modified by the socio-economic context, this ambivalence will also vary depending on the socio-demographic characteristics of respondents such as age, income, gender, ethnicity, and geographic region, reflecting not only ideological differences but also who is systematically harmed by these networks.⁷

Despite these important caveats, the work of anthropologists like Anders and Smith, suggests that in countries like Malawi, ‘the attitude towards the use of the public office for “private” ends is more complicated and multi-layered than the state legislation and regulations seem to suggest’ (Anders 2002, 14). This ambiguity implies that even if aid resources fueled corruption at the local level, local perceptions of corruption may not increase, and may actually decrease, if resources are widely and equitably shared.

Hypotheses

Based on the review of beneficiaries’ perceptions of aid projects discussed above, this study performs statistical tests using geo-coded data on sub-national aid flows as well as survey data on public opinion about perceptions of corruption of government officials, in order to explore the strength and direction of the relationship between aid levels and perceptions and experiences of corruption. I put forward two hypotheses:

H1: Higher levels of aid disbursements in a district will be associated with a greater likelihood of individuals in that district perceiving local leaders as corrupt.

H2: Higher levels of aid disbursements in a district will be associated with a greater likelihood of individuals in that district experiencing corruption in social services.

I have also argued that community members with strong attachments to communitarian norms that oblige sharing and reciprocity may view corrupt practices differently than those who do not hold such views. They may be more judgmental than their non-communitarian neighbors if resources are being misappropriated and kept by a powerful few but they may also be less prone to perceive corruption if misappropriated resources are equally shared within patronage networks. Because theoretically, there is no clear guidance for the direction of the relationship between communitarian values and perceived corruption in aid projects, and because we do not know how equitably misappropriated aid resources are in patronage networks, I test two opposing hypotheses:

H3: Individuals with stronger communitarian beliefs will be more likely to perceive corruption in local leaders as a result of aid allocation than those with weaker beliefs.

H4: Individuals with stronger communitarian beliefs will be less likely to perceive corruption in local leaders as a result of aid allocation than those with weaker beliefs.

Policy relevance

Perceptions of corruption are not a trivial matter. Scholars have theorized that having the perception that political institutions are corrupt degrades overall trust in these institutions (Rothstein 2000). Empirically, increased perception of corruption is correlated with lower confidence in government (Della Porta 2000; Anderson and Tverdova 2003; Bowler and Karp 2004; Redlawsk and McCann 2005; Cho and Kirwin 2007; Tavits 2008; Clausen, Kraay, and Nyiri 2011). It is also correlated with a reduction in belief that democracy is effective (Anderson and Tverdova 2003). Experiences of corruption have also been shown to reduce trust in government officials (Seligson 2002; Anderson and Tverdova 2003; Eek and Rothstein 2009; Morris and Klesner 2010). These effects may be strongest among the poor (Manzetti and Wilson 2007). Activities that significantly increase perceptions or experiences of corruption could, therefore, damage the legitimacy of government, particularly in new and fragile democracies.

To my knowledge, this is one of the first studies to examine the relationship between perceptions of corruption and aid levels at the aggregate sub-national level rather than employing cross-national comparisons or in-depth studies of individual projects. This is also the first study that I am aware of in sub-Saharan Africa that examines the relationship between aid projects and the perceptions of corruption of local leaders, rather than perceptions of the President or national political parties. In the next section, I summarize my empirical approach and afterwards, I report the results of my statistical models and discuss their implications.

Empirical approach

Data sources

This study uses individual- and district-level variables in its statistical models. Individual-level variables include outcome measures of perceptions of corruption as well as socio-demographic control variables that might influence perceptions of local leaders, personal experience of corruption in the use of social services, and the amount of district-level aid. These variables are drawn from public opinion data collected in the second to fifth waves of the Afrobarometer survey (Afrobarometer, 2003, 2005, 2008, 2012, and 2014).

The main explanatory variable is aid disbursement per capita (hereafter ‘aid level’) for the district. Data on aid disbursements are drawn from the AidData ‘Malawi Aid Management Platform’ dataset⁸ (Peratsakis et al. 2012). It contains geographic information for all external aid reported to the Malawi Ministry of Finance from 1997 to 2011.⁹

Main explanatory variable

The amount of aid that a district receives is measured as the mean USD value of aid disbursements in a district for four periods that roughly coincide with each Afrobarometer survey round: 2000–2002, 2003–2005, 2006–2008 and 2009–2011. This figure is then divided by each project's duration and each district's baseline population in 1998 and then logged to address the skewness of the variable.

Outcome variables

The main outcome variable is the perception of corruption of local leaders. The measure is a recoded categorical variable, based on responses to the survey that assesses whether respondents thought that none, some, or all Local Assembly members,¹⁰ local government officials, religious leaders, traditional leaders, NGO or CBO leaders, and local service delivery workers were ‘involved in corruption.’ Responses are coded into a high or low category.¹¹

The other outcome is a variable reporting whether or not the survey respondent has had an experience paying a bribe in the last year for obtaining a social service (water or sanitation services, treatment at local health facilities, and school placement), obtaining official documents, or for avoiding problems with the police. Since evaluations of corruption in local leaders are heavily influenced by actual experiences of corruption, I also include bribery as a control variable in the perception models.

Control variables

I control for the fact that overall perceptions of government corruption might drive perceptions of local leader corruption by including a variable on the level of perceived Presidential corruption in all models. I also include a standard battery of demographic controls, such as gender, age, urban residency, and educational attainment, in all models to mitigate possible confounding between these variables, aid allocation levels, and corruption evaluations. All models include an index of lived poverty because personal economic circumstances might influence feelings about the government as well as the probability of receiving aid. Also included are levels of political interest, media exposure, and closeness with the President's political party, which could shape awareness of, or tolerance for corruption.

I include a measure of group membership, a dichotomous variable that is created by combining responses for two questions asking whether the respondent is a member of a religious group, voluntary association, or a community group. Contact with local patrons is a recoded dichotomous measure of whether or not the respondent has contacted a local ‘influential person’ in the past year with a problem. The local leaders used to construct the contact measure are contact with local councilors, NGO/CBO leaders, traditional and religious leaders. Strength of communitarian beliefs is measured by how strongly respondents say that they agree to the first of these two statements: ‘Once in office, leaders are obliged to help their home community.’ vs. ‘since leaders represent everyone, they should not favor their own family or group.’

Until the 2009 election, voting behavior and perceptions of government performance in Malawi followed ethnic and regional lines closely. I therefore include in most models, indicators for the four largest ethnic groups consistently measured by the survey – Yao, Lomwe, Chewa, and Tumbuka – as well as a residual ‘other’ category.

Finally, I include indicators of existing public goods provision and overall potential for socio-economic conflict within a district, as these factors could confound relationships between aid levels and corruption (e.g. areas that have high social tensions may have diminished ability to attract aid and may be more prone to perceive corruption in local leaders). For public goods provision, I use a measure of public service availability index for the respondent's census enumeration area. The potential for conflict is operationalized by the average ethnic fragmentation in the district and respondents’ perceptions of ethnic discrimination.¹² A sense of ethnic discrimination is measured as the proportion of respondents who report that their ethnic group is ‘never’ treated unfairly (ethnic group treated fairly). Non-responses are coded as an undecided third category. Also included in the models is a measure of respondents’ ethnic attachment.

I include the district-level gini coefficient for income in my models because inequality, interpersonal trust, perceptions of corruption, and income inequality are hypothesized to be linked. A list of study variables and the source from where they are obtained are reported in Table 1 below.

Table 1. Variables used in the analysis.

	Variable	Mean/Percent	Std Deviation	Source
Perceptions of Corruption	High	79%		Afrobarometer Rounds 2–5
District Level Aid (Average per Survey Round)	Aid Disbursements Per Capita	2.9	6.68	Malawi Aid Management Platform
Demographics	(Age) 18–30 years old	46%		Afrobarometer Rounds 2–5
	(Age) 31–40 years old	27%
	(Age) 41 and older	27%
	Male Gender	50%
	Secondary Education or More	20%
	Urban Residence	16%
	Lived Poverty Index	8.76	5.1
Corruption Experiences & Perceptions	Perceive President as Corrupt	63%
	Have Personal Experience of Corruption	11%
Political Attitudes and Partisanship	Political Interest (High)	65%
	Voted for the President's Party	32%
	Media Exposure (High/Medium/Low)	71%
Potential for Conflict	Ethnic Group Treated Fairly (Yes)	53%
	Ethnic Group Treated Fairly (No)	40%
	Ethnic Group Treated Fairly (Undecided)	7%
	Ethnic Fragmentation in the District	0.190	0.110
Ethnicity	Chewa	38%
	Tumbuka	10%
	Yao	14%
	Lomwe	14%
	Other	25%
Membership in Patron or Ethnic Networks	Contact with Local Patrons	37%
	Communitarian Beliefs (High)	43%
	Member of Group	68%
	Ethnic Attachment (High)	48%
Development	Service Availability Index	2.95	1.950	Afrobarometer Rounds 2–5
	Income Inequality	0.370	0.0500	Atlas of Social Statistics

Notes: N = 5860 respondents; taken from the Table 2, column 2, Heckman probit selection model estimation sample.

Percentages for some categories may add up to more than 100% due to rounding.

Regression models

The main outcomes examined in this paper are the answers that survey respondents gave to questions regarding the degree of corruption that they perceived in their local leaders, and whether or not they had recently paid a bribe to receive a social service.

Estimating the impact of aid levels on these outcomes is complicated by two limitations of the data and study design. First, Afrobarometer data are cross-sectional, and may, therefore, be prone to selection bias. Second, there could be endogeneity in my models if aid flows disproportionately to areas that are predisposed to support the government, or to areas that are relatively less corrupt. To correct for these potential weaknesses I report two alternatives to my initial, basic, logistic regression models, both of which use a two-stage approach.

To correct for potential sample selection bias I use Stata's ‘heckprob’ command to fit a Heckman maximum-likelihood probit model with sample selection. The Heckman approach allows unobserved factors that might influence the probability of agreeing to participate in the survey to be related to unobserved factors that might affect perceptions and experiences of corruption.

As a second, alternative approach, I use a two-stage regression model with instrumental variables that could be related to past aid levels but that should not be strongly associated with current perceptions or experiences of corruption. These external instruments are the average road density (m/km²)¹³ in a district in 1998, taken from a 2002 International Food Policy Research Institute (IFPRI) poverty mapping dataset and the baseline 1998 total population in a district from the same source (Benson et al. 2002). These instruments are expected to be positively associated with baseline aid levels because they reflect the geographical convenience of an area to donors and NGOs and the ease of reaching beneficiaries, both factors that should be positively associated with the initial allocation of aid resources to an area (Brass 2012; Briggs 2018). Because I am unsure of the validity of these instruments, I also attempt to deal with the potential endogeneity of aid levels using Lewbel’s (2012) instrumental variable method. Lewbel's method creates an internally generated instrument from the heteroscedasticity present in the data. This internal instrument can be used on its own in a two-stage model, or combined with external instruments whose validity is suspect, in order to strengthen them. I run external and internal instrumental variable models using Stata's ‘ivreg2’ and ‘ivreg2h’ commands respectively.

All models reported (logistic regression, Heckman probit, and instrumental variable models) include the demographic and socio-economic covariates discussed in the previous section, fixed effects for administrative district (N = 28), and fixed effects for survey round (N = 4). Separate models are executed that contain aid interactions with communitarian variables to test whether or not corruption perceptions differ by strength of communitarian beliefs.

Model performance

Both Heckman and external instrumental variable models perform well. In the Heckman selection model, most of the variables in first-stage selection regressions are statistically significant and, based on the results of the likelihood-ratio test I can reject the null hypothesis that the first and second stage of the models are independent. This suggests the selection model might have more consistent and less biased results than a simple probit or logit model.

The external instruments used in the standard instrumental variable model are adequate. In the first-stage equation, the road density measure is not significant for either the perception or experience of corruption models, however, the population variable is significant in both, at the 5% and 1% level respectively. Post-estimation tests of the instruments suggest that the models are statistically valid as the instruments perform well on tests of weak identification and of over- and under-identification (test results are reported in Tables 2 and 3).

The internally generated instrument models on the other hand, perform poorly. The method depends on two important modeling assumptions. The first is heteroscedasticity, as the method constructs potentially valid instruments by multiplying the heteroscedastic residuals from the first-stage regressions with the model's exogenous variables. Breusch–Pagan tests suggest that there is heteroscedasticity in our initial models, so this assumption is met. However, the second assumption, that the squared residuals of the regression are correlated with the dependent variable is not met, and when I evaluate the generated instruments I find that the model fails the Hansen's J over-identification test that would indicate that the internal instruments are exogenous. Because post-estimation tests suggest that the generated instruments may be weak, I do not use them to augment the external instruments but rather report them separately.

Regression results

Regression model results are provided in Tables 2–4 below. I report the average marginal effects of the study variables on the probability of experiencing corruption or of perceiving local leaders as corrupt for ease of interpretation.

Table 2. Aid and perception of local corruption.

	Model 1 (Logistic regression)	Model 2 (Heckman probit)	Model 3 (External instruments)	Model 4 (Generated instruments)
Main Explanatory Variables	Second Stage
Mean Aid Per Capita (log)	0.0174	0.009	−0.246	0.009
	(0.311)	(0.169)	(0.363)	(0.0421)
Perception Presidential Corruption	2.220**	1.213***	0.355**	0.357**
	(0.109)	(0.073)	(0.0172)	(0.017)
Experience Paying Bribe	0.652**	0.337***	0.0733**	0.070**
	(0.173)	(0.088)	(0.0167)	(0.016)
Strength of Ethnic Attachment	0.174^^	0.078	0.0230^^	0.0215
	(0.100)	(0.054)	(0.0136)	(0.013)
Communitarian Feelings	0.0843	0.073	0.0124	0.0146
	(0.148)	(0.082)	(0.0182)	(0.018)
Level of Political Interest	0.0173	−0.004	0.00310	0.003
	(0.105)	(0.059)	(0.0134)	(0.013)
Media Exposure	0.143	0.081	0.0191	0.020
	(0.0933)	(0.051)	(0.0144)	(0.014)
Ethnic Group Treated Fairly (yes)	−0.316**	−0.171**	−0.0440**	−0.044**
	(0.0999)	(0.055)	(0.0138)	(0.014)
Ethnic Group Treated Fairly (undecided)	−0.212	−0.092	−0.0272	−0.024
	(0.191)	(0.103)	(0.0269)	(0.026)
Ethnic Fragmentation	−0.103	−0.113	−0.0151	−0.009
	(0.488)	(0.254)	(0.0567)	(0.056)
Gini Coefficient	0.913	0.393	0.0719	0.096
	(2.390)	(1.233)	(0.299)	(0.294)
Demographics	First, Selection Stage
31–40 Age Group	−0.128	−0.113*	−0.0114	−0.015
	(0.115)	(0.049)	(0.0154)	(0.014)
41 & Older Age Group	−0.215^^	−0.245***	−0.0250	−0.027^^
	(0.116)	(0.049)	(0.0154)	(0.015)
Male Gender	−0.0465	0.303***	−0.00663	−0.007
	(0.0773)	(0.039)	(0.0122)	(0.012)
Educational Attainment	0.541**	0.311***	0.0679**	0.065**
	(0.140)	(0.071)	(0.0166)	(0.016)
Urban Residency	−0.0977	0.073	−0.00167	−0.007
	(0.193)	(0.078)	(0.0325)	(0.031)
Index of Lived Poverty	0.0233	0.005	0.000158	0.0002
	(0.0295)	(0.005)		(0.001)
Ethnicity
Yao	0.0478	0.003	0.00226	0.008
	(0.187)	(0.068)	(0.0241)	(0.023)
Lomwe	0.00344	0.056	−0.00243	0.002
	(0.186)	(0.078)	(0.0244)	(0.023)
Chewa	−0.0229	0.015	−0.00623	0.000
	(0.158)	(0.066)	(0.0216)	(0.019)
Tumbuka	0.274	0.242**	0.0410	0.043
	(0.243)	(0.084)	(0.0328)	(0.033)
Membership in Patron Networks
Patron Contact	0.0234	0.283***	0.00303	0.004
	(0.0920)	(0.050)	(0.0128)	(0.013)
Group Membership	−0.158	0.100*	−0.0206	−0.020
	(0.106)	(0.048)	(0.0138)	(0.014)
Closeness to President’s Party	0.0178	0.421***	0.00687	0.004
	(0.105)	(0.053)	(0.0138)	(0.013)
District Level Development
Service Availability Index	0.0233	0.045**	0.00158	0.003
	(0.0295)	(0.014)	(0.00397)	(0.004)
Instrumental Variable Post-estimation Test Results
Kleibergen-Paap Wald F statistic			22.053	244.058
Weak identification test (Cragg-Donald Wald F statistic)			32.449	244.058
Under-identification test (Kleibergen-Paap rk LM statistic)			44.502***	1464.750***
Over-identification test (Hansen J statistic)			0.859	75.318*
Stock-Yogo weak ID test critical values: 10% IV size			19.93	11.09
Heckman Probit Post-estimation Test Results
Number of censored observations		1040
Number of uncensored observations		4820
Likelihood-ratio test (Independent Equations) (rho = 0):		χ^2  = 5.96*
Observations	4820	5860	5860	4820

Notes: Marginal effects reported; robust standard errors are reported in parentheses for Models 1 and 2 and survey adjusted standard errors are reported in for Models 3 and 4).

The table reports the marginal effects of each variable on the probability of a survey respondent reporting high corruption in local leaders. In addition to the variables listed in the table, models also include dummy variables for district and survey round.

^{^}p < 0.10, *p < 0.05, **p < 0.01, ***p < 0.001.

Table 3. Aid and experience of bribery.

Aid and experience of bribery	Model 1 (Logistic regression)	Model 2 (Heckman probit)	Model 3 (External instruments)	Model 4 (Generated instruments)
Aid Per Capita (log)	0.684^^ (0.352)	0.386* (0.189)	0.619** (0.233)	0.065* (0.032)
Instrumental Variable Post-estimation Test Results
Kleibergen-Paap Wald F statistic			48.70	292.41
Weak identification test (Cragg-Donald Wald F statistic)			68.013	244.058
Under-identification test (Kleibergen-Paap rk LM statistic)			97.058***	1801.55***
Over-identification test (Hansen J statistic)			1.261	82.349**
Stock-Yogo weak ID test critical values: 10% IV size			19.93	11.09
Heckman Probit Post-estimation Test Results
Number of censored observations		222
Number of uncensored observations		5,596
Likelihood-ratio test (Independent Equations) (rho = 0):		χ^2  = 3.80^^
Observations		5818	5810	5810
Alternative Aid Specifications^a
Number of Projects Per Capita (1000)	0.112** (0.0380)	0.443** (0.103)	1.893* (0.779)	0.849** (0.202)
Relative Number of Projects Per Capita (log)	4.368** (1.161)	2.402** (0.611)	3.428** (1.228)	4.642** (1.158)
NGO Implemented Aid (log)	−0.0555* (0.0222)	−2.655*** (0.775)	−0.084* (0.034)	−0.449** (0.157)

Notes: Marginal effects reported; robust standard errors are reported in parentheses for Models 1 and 2 and survey adjusted standard errors are reported in for Models 3 and 4).

The table reports the marginal effects of aid on the probability of survey respondents reporting that they have had to pay a bribe in the past year. Included in the models are all of the covariates listed above in Table 2, except bribe payments. This includes fixed effects for district and survey round.

^p < 0.10, *p < 0.05, **p < 0.01, ***p < 0.001.

^aExternal instruments passed tests of weak instruments, under- and over-identification.

Table 4. Aid and perceptions of local corruption: interaction.

Perceived local corruption	External instrument model
Mean Aid Per Capita (log) per district-year	−0.317 (0.446)
Communitarian Values	−0.0434 (0.0901)
Aid & Comm. Values	0.209 (0.316)

Notes: Marginal effects reported; robust standard errors in parentheses.

The table reports the marginal effects of each variable on the probability of a survey respondent reporting high corruption in local leaders. Included in the models are all of the covariates listed above in Table 2 including fixed effects for district and survey round. External instruments passed tests of weak instruments, under- and over-identification.

^{^}p < 0.10, *p < 0.05, **p < 0.01, ***p < 0.001, N = 4820.

Hypotheses 1: perceptions of corruption

The findings offer no support for the hypothesis that high aid levels lead citizens to view local leaders as corrupt. The relationship between aid levels and perceptions of local corruption does not reach statistical significance in my main models (see Table 2).¹⁴ In robustness tests that look at the number of aid projects in a district rather than the dollar amount of aid, the coefficient on the aid variable is negative, suggesting that increased aid levels might be associated with decreased perceptions of corruption (marginal effect = −0.033, p < 0.05). This observation is contradictory to expectations. The negative trend persists when we use alternative outcome measures of perceived corruption such as perception of Presidential corruption, and perception of corruption in all government agencies. However, higher levels of aid are not associated with better perceptions of government performance.

The factors that have the strongest and most consistent relationship with perceptions of local leader corruption are beliefs about Presidential corruption and having personally experienced corruption by being solicited to pay a bribe. The strength of the Presidential corruption variable supports the idea that perceptions about local leaders are heavily influenced by feelings about the government in general. However, closeness to the President's party has no significant impact on perceptions of corruption in these models.

Sensitivity analysis

The fact that the composition of the local leader corruption outcome measure differs between survey rounds may be of some concern, because we might be measuring slightly different outcomes in each round of the survey. I have tested the sensitivity of the results to the use of different local leaders for the local corruption outcome variable. The estimates are unstable due to large amounts of missing data (>60% of observations missing), and must be interpreted with caution, but overall, I find a non-significant trend towards a negative association between aid levels and perceptions of corruption for local leaders, particularly for religious, traditional, and NGO leaders. The significant exception is local government officials, for whom aid levels are associated with an increase in perceptions of corruption (marginal effect = 0.220, p < 0.001 in the external instrument model).

To examine the discrepancy between the effect of aid on local governmental and non-governmental leaders, I grouped aid into NGO- and government-implemented categories based on project descriptions and titles in the AidData database, and examined the relationship between NGO aid levels and perceptions. I find NGO-implemented aid has negative associations with corruption perceptions, that in some models reach significance (marginal effect = −0.022, p < 0.05 in the Heckman selection model; marginal effect = −.0057, p < 0.05 in the internal instrument model; marginal effect = −2.855, p < 0.10 in the logistic regression model).

One could be concerned that the inclusion of the experience of paying a bribe variable in the perception models is ‘over controlling’ of perceptions of corruption. By holding bribery fixed, I am only measuring the change in the perceptions of corruption that are not based on the experiences of paying bribes. I run models without the bribe measure to address this issue. This modification does not change the magnitude, direction, or significance of the estimated impact of aid level on corruption perceptions, nor does it change the precision of these estimates. Overall perceptions of corruption continue to show no statistically significant association with overall aid levels.

Hypothesis 2: experiences of corruption

When I run models with bribe payment as the outcome, I find that local aid levels are significantly and positively associated with having experiences of paying a bribe for local services (see Table 3). This association occurs regardless of whether aid is measured in dollar amounts or project numbers. A 1% increase in aid levels is associated with a 0.07% to 0.62% increase in the probability of being solicited to pay a bribe. However, the subset of NGO-implemented aid displays the opposite relationship and is negatively associated with bribe solicitation (see Table 3).

The positive association between aid levels and experiences of bribe solicitation could result from aid flowing to more corrupt areas or from aid funding generating opportunities and incentives for misappropriation of resources in communities. The greater experience of bribe solicitation in high aid areas does not translate into higher perceptions of corruption in those areas, although with regards to NGO-implemented aid, lower bribe solicitation and lower corruption perceptions do coincide. In addition, we do not find that the experience of bribe solicitation modifies the relationship between aid and corruption perceptions.

Hypotheses 3 and 4: aid, communitarian beliefs, and corruption perceptions

My findings offer no support for the proposition that those who hold communitarian feelings view the corruption resulting from aid projects differently than those who place less value on sharing and equity. Those who hold communitarian feelings are neither more nor less likely to perceive corruption in their local leaders for a given level of aid regardless of how aid is measured (see Table 4).

Discussion

The study finds no consistent, significant positive relationship between aid levels and citizen perceptions of corruption among local leaders. To the extent that aid is a statistically significant factor in the models tested, it usually displays a negative relationship with corruption perceptions. In other words, higher levels of aid activity, particularly NGO-implemented activity, may be associated with reduced perceptions of corruption. This unexpected result may be due to the perception of the provision of aid resources as a sign of local government fairness and effectiveness. If this is true, the actual level of misappropriation in aid projects may be less important in driving perceptions of corruption than the symbolic presence of these projects in the community.

Despite the absence of a consistent association with perceived corruption, aid levels were significantly and positively associated with more reports of experienced corruption in the form of bribe solicitation. This discrepancy is in keeping with other studies that have found that the perception and true prevalence of corruption may be only loosely associated (Seligson 2002; Seligson 2006; Treisman 2007; Razafindrakoto and Roubaud 2010; Rose and Mishler 2010). In my study, this weak link between perceptions and experience could be explained if perceptions of corruption are an indicator of general perceptions of overall government performance rather than an indicator of actual corruption in the community.¹⁵ Higher-level concerns about government performance and government corruption might be unaffected by the relatively minor resource misappropriation that aid projects produce at the local level.

Another possible explanation for the trend in the data towards a negative trend in the relationship between aid projects and corruption perceptions (and the outlying significant association between aid levels and negative corruption perceptions for local government officials) has to do with differences in corruption between government-implemented and NGO-implemented projects. My bribery findings indicate that corruption in NGO-implemented services is less severe than corruption in government services. The fact that so much aid in Malawi is channeled through NGOs could dampen the association between aid and poor corruption perceptions.

Another possible explanation for the observed negative association between the number of aid projects in a district and perceived corruption, (and for the observation that aid activity might increase the prevalence of corruption but not the perception of its prevalence) might be that misappropriated aid resources are widely distributed and shared within patron-client networks (katangale) so that citizens do not associate aid projects with ‘corruption’ but as rather its opposite: a morally sanctioned, reciprocal sharing of resources. This could also explain why only local government officials, in contrast to other, more socially embedded and better monitored local leaders, buck the trend for a negative relationship between higher perceived corruption and higher aid levels. This interpretation is called into question by the observation that the indicator of communal solidarity does not significantly affect the association between aid levels and perceptions of corruption.

Finally, the difference between bribe solicitation and perceptions of corruption could be explained if paying small bribes to receive services were not categorized as an exceptional circumstance by individuals but rather as a normal part of doing business and receiving social services. If such solicitation is an entrenched part of Malawi's social service system, it would be logical that such solicitation would increase along with increases in the provision of social services that aid resources bring about. By increasing the availability of social services, aid would create more opportunities for bribery solicitation, particularly in government facilities.¹⁶

Concluding remarks

Studies of the local political impact of individual aid programs are almost unanimous in finding that aid projects support patronage structures and reinforce local power dynamics. The increasing worry voiced by those concerned with aid effectiveness is that these projects might cumulatively retard political development by strengthening local leaders and by reducing incentives for them to act accountably. The results from this study do little to allay these fears. The strong positive association between aid levels and the experience of corruption suggests that aid resources might feed into existing patron-client resource distribution networks and erode support for local government officials. Our findings do, however, support the contention that channeling aid through NGOs might mitigate these negative effects of aid.

In regards to future studies, additional research on sector- and donor-specific differences in the relationship between aid levels and corruption would be valuable, as would the replication of this study in other sub-Saharan African countries for which data on sub-national aid flows are becoming available. More scholarship from Malawians themselves on their perceptions of aid and corruption would also be welcome.

Disclosure statement

No potential conflict of interest was reported by the author.

ORCID

Sahai Burrowes http://orcid.org/0000-0002-4944-402X

Notes

1 Throughout the paper I employ the standard definition of ‘corruption’ to mean the abuse of public office for private gain (World Bank/IMF 2007).

2 This is mostly the case for bilateral donors. Multilateral donors, being somewhat sheltered from these pressures, may have a greater incentive to monitor projects. There is some indication that multilateral aid may have a different impact on corruption than bilateral aid (Charron 2011).

3 They are however at odds with Olken's study (2009) which found that villagers’ perceptions of corruption were highly and positively correlated with a measure of the actual amount of corruption in the project.

4 Such networks are particularly important in times of economic uncertainty when they serve as safety nets and a way to pool resources. Malawi's post-transition period with its rapid urbanization and reoccurring economic crises might be considered a period during which these networks would have been particularly important.

5 The usage of the term katangale may have shifted over time, as, contrary to Anders, I find that in current popular media and online discussions, katangale is always referred to negatively to describe a general state corruption, while kuba is used to describe particular episodes of theft resulting from katangale. The changing usage could also reflect a decreasing tolerance of katangale although it should be noted that while public opinion surveys show that the vast majority of Malawians reject the notion that leaders should ‘help their own community’ the proportion of respondents who agree with that statement has been growing, not declining over time (Afrobarometer 2003, 2005, 2008, 2012, and 2014).

6 Smith's work also suggests that the invocation of norms of reciprocity in resource allocation may be used as a critique of personal enrichment at public expense and of the inequality it breeds. In other words, nuanced and ambivalent attitudes towards katangale may not reflect intransigent, static ‘traditional’ beliefs, but rather dissatisfaction with, and reaction to, the economic displacement brought about by globalization, systematic graft, and political closure.

7 See Yang 2002 for an interesting discussion along these lines regarding guanxi in China.

8 The dataset is publically available at http://www.aiddata.org/content/index/AidData-Raw/geocoded-data

9 Projects before 2000 were not exhaustively catalogued so the database is only considered complete for the 2000–2011 time period.

10 The Local Assembly is Malawi's district-level legislative body.

11 The survey question was ‘How many of the following people do you think are involved in corruption, or haven't you heard enough about them to say:__?’ It is coded as a 0 if the respondent says ‘none’ or ‘some of them’, 1 if the respondent says ‘most of them’, or ‘all of them’. Don't know and non-responses were dropped from the analysis. Responses about the following local patrons are included in the measure: local councilors (all rounds), traditional leaders (rounds 4 and 2 only), religious leaders (round 2 only), NGO or CBO leaders (rounds 2 and 5), and local service providers (round 3 only).

12 Ethnic fractionalization is measured by a Herfindahl-Hirschman index for the number of different ethnic groups in each district in each survey round. I create the variable using Stata's user-generated ‘hhi’ command (Ansari 2012).

13 This is the average meters of road per 100 sq. km of land area weighted by the ‘potential speed on different qualities of road’ and deflated population size.

14 Aid also has no statistically significant impact on perceived Presidential corruption, perceived overall government corruption or on perceptions of local government or Presidential performance.

15 Local government performance evaluations and local leader corruption measures are significantly (although weakly) correlated in this study (r = −0.06, p-value < 0.001).

16 My thanks to Shana Warren for this suggestion.