Political Selection and Bureaucratic Productivity

Economic theory of public bureaucracies as complex organizations predicts that bureaucratic productivity can be shaped by the selection of different types of agents, beyond their incentives. This theory applies to the institutions of local government in the developing world, where nationally appointed bureaucrats and locally elected politicians together manage the implementation of public policies and the delivery of services. Yet, there is no evidence on whether (which) selection traits of these bureaucrats and politicians matter for the productivity of local bureaucracies. This paper addresses the empirical gap by gathering rich data in an institutional context of district governments in Uganda, which is typical of the local state in poor countries. The paper measures traits such as the integrity, altruism, personality, and public service motivation of bureaucrats and politicians. It finds robust evidence that higher integrity among locally elected politicians is associated with substantively better delivery of public health services by district bureaucracies. Together with the theory, this evidence suggests that policy makers seeking to build local state capacity in poor countries should take political selection seriously.


Policy Research Working Paper 8673
Economic theory of public bureaucracies as complex organizations predicts that bureaucratic productivity can be shaped by the selection of different types of agents, beyond their incentives. This theory applies to the institutions of local government in the developing world, where nationally appointed bureaucrats and locally elected politicians together manage the implementation of public policies and the delivery of services. Yet, there is no evidence on whether (which) selection traits of these bureaucrats and politicians matter for the productivity of local bureaucracies. This paper addresses the empirical gap by gathering rich data in an institutional context of district governments in Uganda, which is typical of the local state in poor countries. The paper measures traits such as the integrity, altruism, personality, and public service motivation of bureaucrats and politicians. It finds robust evidence that higher integrity among locally elected politicians is associated with substantively better delivery of public health services by district bureaucracies. Together with the theory, this evidence suggests that policy makers seeking to build local state capacity in poor countries should take political selection seriously. This paper is a product of the Development Research Group, Development Economics. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/research. The authors may be contacted at skhemani@worldbank.org.

Introduction
The impact of anti-poverty and pro-growth policies and spending programs depends upon how well they are implemented or delivered by government bureaucracies. The quintessential "delivery unit" in developing countries consists of local governments helmed by appointed bureaucrats and locally elected politicians. Institutional arrangements in the local state, or the last mile at which delivery and implementation happens across countries such as India, Indonesia, Nigeria, Brazil, Tanzania and Uganda, provide space for locally elected politicians to facilitate and monitor the administration of nationally funded programs (Ahmad andBrosio, 2006 and2009). These local institutions correspond to the setting examined in principal-agent theory of the public sector: multiple principals and agents; and, complexity of tasks and objectives (Tirole, 1994;Dewatripont, Jewett and Tirole, 1999;Dixit, 2002). The theory suggests that productivity in such organizations can be enhanced through organizational design that increases motivation and improves selection of agents, going beyond incentives (Besley and Ghatak, 2005;Acemoglu, Kremer and Mian, 2007;Akerlof, 2017). However, there is little to no empirical evidence on whether selection of agents matters, and furthermore, which traits matter for performance. 1 We provide the rst evidence on how political and bureaucratic selection traits correlate with the productivity of public sector organizations. We measure traits of local politicians and bureaucrats in an institutional context which is typical of the local state in poor countries-district governments in Uganda-and examine whether these measures are robustly and substantively correlated with variation in service delivery by these districts.
Surveys were undertaken during September-December 2015 in 75 of Uganda's 112 districts at that time, 2 the relevant local jurisdiction where both directly elected district politicians and nationally appointed bureaucrats share responsibility for implementing public policies and delivering services. Modules available from the literature to measure integrity, altruism, cognition, risk-aversion, personality traits, and public service motivation, were administered to 1,357 district bureaucrats and 770 district politicians. Administrative data available on the implementation of nationally mandated public health programs serve as a measure of productivity of the district organization. These data reveal that the integrity of politicians is a robust predictor of district productivity in delivering health services. The size of the correlation is substantively large: a 1 standard deviation higher average integrity among local politicians is associated with a 0.2 to 0.4 standard deviation higher measure of performance in health service delivery. In terms of concrete indicators of health coverage, an increase of 1 s.d. in the average integrity of politicians is correlated with a 4.4 percentage point increase in child delivery at government facilities (9% increase at the mean); a 4.7 p.p. increase in share of households with latrines (6.2% at the mean); a 3.8 p.p. increase in share of pregnant women preventively treated for malaria (6.7% at the mean) and a 2.8 p.p. increase in pregnant women having at least 4 antenatal care visits (8.2% at mean).
In contrast to the correlation with political quality, we nd no signi cant correlation of service delivery with measures of electoral competition such as the concentration of vote shares of competing candidates and the margin of victory. At the same time, we nd that political integrity tends to be higher in places where there is greater electoral competition. This is suggestive evidence that competition can play a role through improving selection (such as, by enabling the election of politicians of higher integrity) rather than through incentives alone. Our results are consistent with other research that incentives generated by electoral competition may play an ambiguous role in public service delivery. 3 Our evidence also brings into the fore a hitherto neglected reason why quality (selection) of politicians might matter beyond incentives -because public sector bureaucracies are complex orga- 3 Other research has cast doubt on the role of electoral incentives in improving outcomes. For example, political incentives to seek re-election can distort the allocation of public funds from what a benevolent social planner would choose (Finan and Mazzocco, 2017). Political parties can respond to greater electoral pressures by elding criminals as candidates (Aidt, Golden and Tiwari, 2015), or by increasing vote-buying (Cruz, Keefer and Labonne, 2016;Khemani, 2015), rather than by improving services. Afridi et al (2017) provide both theory and consistent evidence that corruption in developing countries can increase with electoral competition. Consistent with our results, Grossman and Michelitch (2018) nd that Ugandan local politicians do not respond to increased incentives (through a transparency intervention) in competitive constituencies by improving service delivery. nizations whose performance depends upon the interaction between a large number of principals and agents. Stronger electoral incentives of one set of actors -politicians -may not generate the actions that are conducive to organizational or team productivity.
The evidence we contribute on political integrity lends broad support to the theory of public bureaucracies as complex organizations. In district governments in Uganda, as in local government agencies of most poor countries, two types of agents share responsibility for managing a swathe of public spending programs and implementing policies-nationally appointed bureaucrats and locally elected politicians. 4 While the nationally-appointed bureaucrats are expected to bring the technical competence needed for public sector management, the locally elected politicians are supposed to play a role in monitoring and facilitating service delivery, especially by liaising with citizens and communities, increasing public awareness and winning public support. The cadre of politicians whose quality (integrity) is revealed by our data as a signi cant correlate of service delivery-locally elected district councilors in Uganda-are elected from constituencies (sub-counties) at the frontlines of service delivery, where health clinics operate and public health workers perform their tasks. Monitoring these health service delivery units is cited as one of the signicant duties that local councilors are expected to perform (Grossman and Michelitch, 2018).
However, councilors have been described as typically failing to perform their duties of political oversight (Ra er, 2016). Our nding thus suggests that where these local politicians are of better quality (higher integrity), they undertake their monitoring and facilitating roles more e ectively, thereby contributing to improvements in service delivery. While prior work has focused on the policy-making role of more powerful political leaders, at national or regional levels, our work shows that the quality of more local politicians, who wield facilitating and monitoring powers over service delivery, also matters.
This evidence on political integrity is consistent with theory on the importance of political selection. The paper thus contributes to lling a large gap between economic theory and empirical evidence on political selection (Besley, 2006 andDal Bo et al, 2017, provide reviews). The identity and characteristics of leaders can matter because the political institutions through which leaders gain o ce (such as elections) o er only incomplete contracts, and lack of enforceability by third parties (Osborne and Slivinski, 1996;Acemoglu, 2003).
Furthermore, the quality of politicians can matter because, as leaders, they can play a role in shaping social norms and policy preferences (Acemoglu et al, 2015;Bidner and Francois, 2013). Yet, despite strong theoretical reasons to care about political selection, there is little evidence of whether and how it matters. The few empirical studies available on political selection are constrained by not having access to data on typically unobservable dimensions of quality, such as integrity or honesty, and altruism or a sense of civic duty, which may not be captured by available measures such as education (Ferraz and Finan, 2009). 5 Indeed, in our own data we nd only weak and small correlations between education and di cult-to-observe personality and behavioral traits. 6 Theory suggests that self-selection of political candidates on these di cult-to-observe variables, speci cally, integrity and altruism, can play an important role in governance (Bernheim and Kartik, 2014). We nd, in fact, that political integrity matters for service delivery.
We recognize and emphasize throughout the paper that this study can only report robust correlations, not a causal relationship. We cannot identify an exogenous institution that drives variation in political selection. It may be that some omitted variable, such as citizen preferences for public goods, is correlated with better quality citizens entering the political market as candidates, better quality candidates getting elected, and with better service delivery by health workers. We argue that the way to interpret the correlation we nd in the data is arising from the interaction between citizens, local politicians and service providers, as in the framework provided in World Bank (2016a) of inter-dependent principal-agent relationships in government. This interaction yields an equilibrium in which citizen demand, political selection, and behavior of health workers are all endogenously determined. The correlation between political integrity and local state productiv-5 One strand of the empirical literature on political selection has focused on the phenomenon of adverse selection of criminals as political candidates in India (Aidt, Golden and Tiwari, 2015;Asher and Novosad, 2018;George, Gupta and Neggers, 2018). 6 Correlations reported in Table 3. ity is thus driven by variation across the 75 di erent equilibria of district governments.
We argue and provide supporting evidence to interpret the correlation as follows: in districts where higher integrity politicians are endogenously selected, these politicians exert greater e ort towards public goods than rent-extraction; as a result, the local implementation of national health programs improves. Even if omitted variables that capture citizens' political preferences or prevailing social norms are the underlying driver of variation in political selection, the proximate channel of in uence of political integrity on service delivery is through the e ort exerted by selected politicians. We show that the correlation with political integrity holds even after accounting for plausible sources of variation in voter demands (such as, partisan attachment to national leaders, and the presence of local radio), thus suggesting that local politicians have a direct or proximate channel of in uence on service delivery, apart from other channels through which citizens might in uence service delivery without going through politicians.
We further provide qualitative evidence and support from other studies in Uganda of the role of local councilors in monitoring local health facilities, and how they can play these roles in obstructionist, rent-seeking ways (such as shielding absent health workers from disciplinary action, or pilfering funds at the health clinic) versus facilitating better delivery (such as encouraging health workers to show-up, and monitoring that clinic funds are appropriately used). We suggest that the measure of integrity is capturing the proclivity of politicians to exert e ort in one type of action compared to another-whether in the extraction of private rents versus the promotion of broader service delivery. Higher integrity politicians are more motivated to pursue public goods and therefore more e ective in improving public health service delivery.
Nevertheless, omitted variables that explain both integrity and the ability of politicians to in uence service delivery could be biasing the coe cient on integrity that we estimate.
Random assignment of higher integrity politicians to districts may yield di erent estimates of impact on service delivery. However, random assignment of integrity may not be the appropriate way to test theories where political selection arises endogenously. Learning from correlations revealed by the data is valuable to test theories of endogenous relation-ships that arise in equilibrium, and to inform further theoretical development. For example, models of political selection have been quite separate from models of agent selection in complex organizations of the bureaucracy. Our results suggest scope for further developing a theory of how these two di erent types of agents, bureaucrats and politicians, interact in government agencies. Our study provides unique new evidence to understand the relative characteristics of these agents.
We nd systematic di erences across several dimensions of quality between politicians and bureaucrats, and, in which features of quality of these two types of agents matter for service delivery. For example, bureaucrats are more educated than politicians and education of bureaucrats is correlated with service delivery; in contrast, education of politicians is not correlated with service delivery. These di erences in education between bureaucrats and politicians may be viewed as simply con rming what one might have suspected given the local institutional context where the bureaucrats have greater powers and are appointed from a professional national cadre, while the politicians are drawn from local communities and have fewer formal powers. Other di erences, in integrity, altruism, personality, and public service motivation, are areas with fewer priors and more to learn from the data. We nd that politicians report greater altruism when compared to bureaucrats.
However, while altruism among bureaucrats matters for service delivery, it is the integrity of politicians that is associated with improved services (not their reported altruism or public service motivation). And it is political integrity which is in short supply: politicians score signi cantly lower than bureaucrats, and those who win elections tend to score less than contenders who lost.
Our results suggest that variation in personality of politicians plays a smaller role than variation in personality of frontline service providers as found in the work of others (Donato et al, 2017;Callen et al, 2015). Traits like conscientiousness and neuroticism might matter much less for politicians than service providers because the latter are engaged in day-to-day active service delivery. Integrity, on the other hand, has been highlighted in theory as a particularly important characteristic of politicians (Bernheim and Kartik, 2014).
However, while political integrity is revealed by the data as the most robust correlate of service delivery, to the exclusion of available measures of personality, altruism, or public service motivation of politicians, this could be because these other characteristics are harder to measure and what is available has not been fully validated in developing country settings (Laajaj and Macours, 2017). Our contribution is the tip of the iceberg in this regard, and a necessary one to build empirical evidence on political and bureaucratic selection in developing countries. Institutional space for the emergence of local politicians has been spreading across and within countries, even where national politics has veered towards authoritarianism (World Bank, 2016a). More research and policy experimentation are warranted to understand the quality of these emerging political leaders at the local level, and how they can play a role in building state capacity. Available research has focused on understanding the incentives of 7 A notable exception is the work of Ra er (2017) which evaluates the impact of training local politicians in Uganda to play a more e ective role in monitoring the implementation of public investment projects. The results suggest that better trained councilors, who are provided nancial and technical information about the projects being implemented in their constituencies, can become more active in exerting political oversight on the projects, but only in districts where the chairperson does not belong to the national ruling party. Ra er (2017) interprets this pattern as suggestive of weaker political incentives to monitor where the national ruling party is more powerful. Our nding of a negative correlation between voter support for the national political party's presidential candidate and health service delivery is consistent with Ra er's (2017) ndings. Our contribution of examining the role of political selection goes beyond incentives and supports policy directions to encourage better quality citizens to enter local governments as political leaders. politicians to perform, to the neglect of understanding selection. For example, Grossman and Michelitch (2017) is a recent study from Uganda, where research e orts were devoted to evaluating a transparency intervention targeted at strengthening political incentives.
Interestingly, the study nds no e ect on service delivery, and the authors explain this precisely by appealing to the role of multiple actors in service delivery organizations. A few studies are beginning to examine whether transparency interventions can be designed from the start to impact political selection, nding that entry of candidates is quite responsive to such interventions, resulting in greater representation of non-elite citizens (Gulzar and Khan, 2017;Banerjee et al, 2017). Much more work remains to understand whether such transparency interventions result in the selection of particular political traits, whether these traits matter for the performance of public sector organizations, and through what mechanisms.
The rest of the paper is organized as follows. Section 2 describes the institutional context and data gathered in Uganda to examine the role of selection traits of bureaucrats and politicians in the productivity of the district organization. This section presents evidence of systematic di erences between bureaucrats and politicians across several dimensions of selection. Section 3 provides the empirical speci cation to examine the correlation between district productivity and selection traits of bureaucrats and politicians. Section 4 presents the main results of the robust correlates of local productivity in delivering health services. Political integrity emerges as a particularly signi cant and substantial covariate even after accounting for bureaucrat selection, and other socio-economic-political factors of the district. Section 5 examines the correlates of political integrity-what conditions are associated with the selection of better quality politicians. Section 6 discusses the pattern of evidence and provides an interpretation of the mechanisms behind it. Section 7 concludes.

Institutional Context and Data
The district government in Uganda exhibits institutional characteristics that are typical of the local organizations that deliver public services and implement public policies in devel-oping countries. Bureaucrats appointed to the district by national ministries are responsible for managing national scal transfers to deliver public services across sectors-health, education, agriculture, roads, water, environment and natural resources. Following a wave of political decentralization across the developing world in the 1990s, Uganda promulgated The Local Government Act of 1997 to provide space for locally elected politicians in the functioning of the district bureaucracy. District councils, consisting of councilors elected from sub-district constituencies and headed by a directly elected chairperson, have legislative functions over matters devolved to the local council by the national state. Most importantly, these locally elected politicians are expected to perform unique roles of oversight or monitoring the bureaucracy, and of facilitating service delivery, as a liaison between local citizens and the state (Grossman and Michelitch, 2018;Ra er, 2017). An original survey was undertaken between September and December 2015 in 75 of Uganda's 112 districts (at the time) to measure the selection traits of these district bureaucrats and politicians who oversee the implementation of national programs. 8 Figure 1 provides a map of our study area.
In each District, all the following elected politicians were interviewed: the District Chairperson or Municipality Mayor; three District/Municipality Woman Councilors and three District/Municipality Directly Elected Councilors. In addition, all non-elected contenders for the position of District Chairperson were listed and approached for interviews.
In total 770 district politicians and political contenders were interviewed. Among bureaucrats, all District Chief Administrative O cers (CAO), deputy CAO and the Resident District Commissioners (RDC) were interviewed. An extensive roster of senior o cers were also included on the sample to be interviewed 9 ; whenever they refused to be interviewed or the position was vacant, they were replaced by other members of the same department.
A total number of 1,357 bureaucrats are included in the nal sample. Our survey respondents thus comprise key o cials in leadership positions in local government, with two distinct lines of selection: (1) local politicians who contest elections and are voted in/out by locally-residing citizens; (2) bureaucrats who are appointed by national political leaders and technical ministries. We asked a series of questions in our survey about who has leadership powers in local government, and found evidence of an informal sharing of power between locally elected politicians and bureaucrats appointed by higher-tier politicians. For example, we asked: "Who in your district/municipality has the greatest ability to get civil servants to follow their orders?" Among the appointed bureaucrats and technocrats responding to this question, 93 percent indicated that the Chief Administrator O cer (CAO), who is appointed by national politicians, has the greatest authority. However, among the politicians responding to this question, a lower proportion, 70 percent, indicated the CAO, and 24 percent indicated that the directly elected District Chairperson or Municipality Mayor has the most authority. Local journalists-who head the radio stations which comprise the main news media serving these districts-were even more split in their responses: 43 percent indicated the CAO, 24 percent indicated the directly elected local politicians, and 21 percent indicated the Resident District Commissioners, who are another cadre of senior bureaucrats appointed by national politicians. This setting thus corresponds to the special features of public sector organizations that was discussed in Dixit (2002)-multiple stake-holders (principals) who together manage a complex set of policy implementation and service delivery activities. The key distinction between the two types of local leaders is that one type, the politician, is selected locally from among the residents of the local government through the electoral process, while the other, the bureaucrat, consist of career technocrats selected by national political and bureaucratic leaders from a national pool of candidates. 10 Survey modules were developed on the basis of the available literature to measure the integrity, altruism, cognition, risk-aversion, personality traits, and public service motiva-tion of these bureaucrats and politicians, and are discussed in the section below. To our knowledge, this is the most comprehensive data gathered on a large number of key actors interacting in the complex organization of the public sector.

Selection traits of bureaucrats and politicians
Detailed modules measuring individual characteristics and behavioral traits were administered for both politicians and bureaucrats. Basic demographic characteristics include gender, age and educational attainment, as well as household-level characteristics such as family size and asset ownership. Other than basic demographic characteristics, our survey includes rich information about individuals' integrity, cognitive ability, risk preferences, non-cognitive traits and motivation to work in the public sector. Detailed information about each of these variables and the construction of aggregate measures are provided in Annex 9.1 ; here we provide a brief description of these behavioral measures.
The integrity of respondents was assessed through the Moral Disengagement Measure, developed by Moore et al. (2012). The Integrity Index is a z-score using responses from eight sub-components. Taken together, the eight questions presented in Annex 9.4 have been shown to be correlated with unethical behavior in lab experiments. Individuals are asked to rate their agreement with the statement on a scale of 1 (Strongly Disagree) to 5 (Strongly Agree). We recode the variables such that higher values correspond to disagreement with unethical behavior, then proceed to take a simple average of the eight questions and normalize it to obtain the Integrity Index.
In assessing non-cognitive ability, the survey follows the World Bank STEP (Skills Towards Employability and Productivity) methodology. It includes not only the popular Big Five Index -a set of 15 questions used to assess ve broad domains of personality -, but also the more speci c dimensions of Grit, Decision-Making and Hostility Bias. The motivation to work in the public sector is measured using Perry's PSM Index, composed of 36 questions measuring motivation in six di erent dimensions -Attraction to Policy Making, Commitment to Public Interest, Social Justice, Civic Duty, Compassion and Self-Sacri ce 11 . Risk Preferences are assessed using a simple, non-incentivized risk game, detailed in the Annex 9.7. Similarly, Altruism is measured by a hypothetical game in which the individual is given UGX 50,000 12 and has to decide how much of it to share with an anonymous individual.
To assess cognitive ability, the survey uses a digit span memory test, in which an increasingly longer string of numbers is read to the respondent, who is in turn asked to repeat it. We use the number of digits the respondent can correctly recall as a proxy for cognitive ability. 13 . We complement this measure, following Dal Bo, Finan and Rossi (2013), with an indicator assessing whether the respondent chooses a dominated option in the risk game described above.
Tables 1 and 2 present descriptive statistics on the survey sample, highlighting di erences in means between politicians (including non-elected contenders) and bureaucrats. Table 1 presents basic socio-demographics characteristics and con rms the intuition that politicians are very di erent from bureaucrats in basic observable characteristics. Politicians are much more likely to be females, to have been born in the same District where they currently work 14 and also to come from less wealthy households (they have larger families and hold less assets). The most striking distinction between bureaucrats and politicians is in educational attainment: whereas half of politicians have not completed college, 96 percent of bureaucrats did, including half with postgraduate degrees. These numbers are broadly consistent with Ra er's (2017) data on subcounty politicians and o cials in Uganda, and paint broadly the same scenario of policy environment: local politicians oversee project execution led by highly educated and centrally appointed bureaucrats. Table 2 presents descriptive summaries of selection traits. Panel A reports, for each of the components on the Integrity module, the share of individuals who agree or strongly agree with each statement. For all components a larger share of politicians than bureaucrats agree with statements such as "People shouldn't be held accountable for doing ques- Finan and Rossi (2013) and Callen et al. (2015) 12 Approximately USD 15, at the conversion rate of USD 1 = UGX 3,400 (average during the survey period). 13 This is the same methodology used in Weaver (2016). 14 Which is not surprising given that bureaucrats are centrally appointed. fewer numbers than bureaucrats, and are also 3 p.p. more likely to choose a dominated option on a simple risk game (choosing a UGX 1 million reward instead of a gamble of equal probability between UGX 1 million and UGX 2 million) 16 . Regarding altruism, politicians donate a slightly larger fraction of a hypothetical prize on average, and also make less risky choices in lottery games. 17 15 One advantage of using the same modules as other surveys in the literature is comparing the results obtained in Uganda with other settings. This paper uses the same modules for measurement of PSM and Big-Five as Dal Bo, Finan and Rossi (2013) in Mexico, so we can actually compare some of the descriptive statistics to their Table II. Our mean values for Big Five measures are across the board lower, indicating a "worse" pool of individuals in terms of these non-cognitive traits. For PSM, however, the scores are pretty similar. 16 The share of individuals who make a mistake, for both politicians and bureaucrats, around 10 percent in the Uganda sample, stand in stark contrast to the results of Dal Bo, Finan and Rossi (2013) where almost 40 percent of individuals who were being considered for the position of social workers in Mexico made this error.
17 While all gures discussed in these tables refer to di erences in means, Figures 3 and 4 presents the CDF of traits for each group and Kolmogorov-Smirnov tests reject equality of distributions.
While Table 2 suggests politicians have very distinct behavioral traits from bureaucrats, it is possible that all these di erences disappear when we compare individuals with similar observable characteristics. Is that the case, or do individuals who select to be politicians fundamentally di er from those at bureaucratic positions once we control for observable traits such as gender, education and age? Figure 2 presents the average di erence in traits between politicians and bureaucrats, once we account for a range of observable individual characteristics. Most of the previous results stand: politicians still perform worse in Integrity and Other Non-Cognitive measures (Grit, Decision Making and Hostility Bias), but better in the PSM, Big Five and Altruism. The main distinction from the previous ndings is that di erences in cognitive attainment become very small and statistically insigni cant, suggesting that any di erences between the two groups are simply mediated by di erential educational attainment. 18

Consistency of individuals' traits measures
One possible caveat to this discussion is the reliability of these behavioral measures, which are self-reported. Whereas we cannot rule out entirely that desirability bias dominates our measurement, we provide some evidence they seem to capture important dimensions of personality.
First, as discussed above, there's quite a lot of variation in answers. Even for questions that might induce a strong desire to conform, such as "Is it OK to take credit for others' ideas?", 20 percent of respondents answer they disagree and only 20 percent assert they strongly agree. Furthermore, we identify large and systematic di erences between answers of politicians and bureaucrats. Second, Table 3 Table 4, top panel, presents descriptive statistics on the quality of health services across districts. Whereas there is little variation in some of the individual indicators (the median coverage of DPT3 vaccination is 98 percent), other indicators vary widely between Districts: the average share of women having 4 or more ante natal care visits is only 34 percent (s.d.: 10.9 ) and the average share of pregnant women receiving preventive malaria treatment in pregnancy (IPT2) is 53 percent (s.d.: 11.6). The DLT total score is a weighted average of 14 quality care and management practice indicators ranging from 0 to 100 -the average District scores 73 points. In the analysis below we will often use a normalized version of the total score, so that changes can be interpreted as standard deviation variation in the score (7.25 points).
These health service delivery indicators are the best measures we have available in our study context to capture the productivity of the district bureaucracy as a complex organization. Rasul and Rogger (2016) and Rasul, Rogger and Williams (2017) review the available measures of productivity of national bureaucracies and focus on completion rates of projects undertaken. The advantage of the setting we examine-of district-level bureaucracies-is that productivity can be measured in terms of services that are directly linked to development outcomes, such as health in poor communities.

Empirical Speci cation
The empirical analysis in this paper is not positioned to causally identify the e ect of any variable on health service delivery performance. Rather, we examine what are the robust correlates, if any, of variation in performance across districts using all available data on a variety of potentially signi cant correlates for this purpose. The main question guiding the analysis is whether (which) selection traits of bureaucrats and politicians are significant in explaining district performance. In order to control for the role of other district socio-economic and political characteristics in driving any correlation, by simultaneously in uencing both selection traits and health service delivery, we use other District level administrative data on these district characteristics.
Speci cally, we estimate variants of the following equation: where Y ij will often be the standardized health score in the District League Table for district i in region j. In some instances, Y ij will instead be speci c components of the DLT score. The vectorX i includes our main variables of interest: the average trait of politicians and/or bureaucrats in District i; di erent speci cations will use di erent subsets of traits.
Z i is a vector including a series of socio-economic and political characteristics of districts and γ j allows for region-speci c intercepts.
The variables contained in the vector Z i come from administrative and survey data. The 2014 Census provides information on Districts' population and urbanization rates, while poverty rates are estimates for 2013 (World Bank, 2016b) 21 . We complement these socioeconomic data with average students' score at the UCE in 2011, a nationally administered exam students take at the end of lower secondary school, as the best available proxy for education in the population. 22 Our survey also gathered data on the population of radio stations operating in Uganda at the time of the survey, from which we derive a variable of the number of radio stations headquartered in a district. Access to radio, and to media in a more general sense, has been shown to impact the public scrutiny of local politicians (Ferraz and Finan, 2008) and a ect citizens' investment in health and education (Khemani and Keefer, 2015;Khemani and Keefer, 2014). We also consider that the existence of local radio stations might proxy for other unobservables such as the dynamism of local markets.
In addition, we compile data on local vote shares from the 2011 elections, the last one before our survey. The main variables are the vote share of the presidential candidate of the NRM, the national ruling party (labeled "Voter Attachment to National Party"), at the district level, and voting Her ndhal Indices 23 for district politicians 24 .
The lower panel in Table 4 presents descriptive statistics of administrative data for the 75 Districts included in the survey. The median District in our sample has approximately 250,000 people and is overwhelmingly rural 25 . Poverty rates, at 23 percent in the median District, are only slightly higher than the national rate of 20 percent in 2013. The numbers also show the subdued level of political competition in presidential elections: the candidate 21 These are small areas estimates produced by the World Bank, using data from the 2013 UNHS and 2014 Census data. For details on methodology see for example Ebers, Lanjow and Lanjow (2002). 22 Note that higher values of the UCE score are indicate of lower education performance, as described in the data Annex. We include 2011 UCE performance since those cannot be in uenced by current politicians, who were elected in that year. 23 The Her ndahl Index is most commonly used as measure of rm competition in a given market, but has been widely used to measure political competition . The Her ndahl Index is obtained, at the electoral area level, by the sum of the squares of candidates' votes. Higher values are associated with a less competitive environment (the index ranges from 1/N to 1, where N is the number of candidates). 24 We compute all the results at the electoral area level and then obtain the district-level average across all areas. 25 According to the 2014 Census only 16 percent of the population in Uganda lived in urban areas.
of the ruling party won 70 percent of the popular vote for president in the median district during the 2011 presidential election.

Results: correlates of district productivity
This section presents the main results of the paper, assessing the correlation of politicians' and bureaucrats' selection traits and service delivery. We start by presenting raw correlations between actors' selection traits and districts' health outcomes. Here we document that integrity is the only trait of politicians' that is correlated with health outcomes, while among bureaucrats' both risk-aversion and altruism show positive correlations with health quality.
Our main result is to assess whether these correlations hold in multivariate regressions.
Because of the problem of having a small number of observations (75 districts) relative to a large number of potential regressors (40 in total, counting all selection traits and district socio-economic and electoral characteristics on which we have data available), we face the issue of model selection -should we just include all covariates, at the cost of having very noisy parameter estimates, or should we try to restrict the model to covariates we a priori think are "better"? The available economic theory provides little guidance on this.
Our reading of the literature suggests that a machine-learning process such as LASSO has become the methodology of choice in such situations, to allow the data to reveal which variables have greater predictive power. We therefore use a model that includes variables chosen by LASSO -which drops variables that explain little of the variation in the dataas our main speci cation. We think this strategy should make our readers less worried that we are "cherry-picking" a model that makes some results look better. We also present in the robustness section a range of OLS estimations with an exhaustive list of controls.
These estimates suggest that politicians' integrity is the most robust predictor of variation in health outcomes among all selection traits of policy actors.
We then proceed to perform a series of additional robustness checks on our main result.
Having established the role of politicians' integrity as an explanation for overall health service quality, we show that the correlation holds not only for the aggregate score but also for a majority of individual sub-components of the index. The results become stronger if we use the rst principal component of individual health indicators.
Finally, we provide evidence consistent with the mechanism that higher political integrity in uences bureaucratic productivity through improved oversight and facilitation by district councilors who are elected at the frontlines of health service delivery. Among the three types of politicians surveyed -local councillors, District Chairperson and nonelected Contenders -, it is the integrity of councillors, who are supposed to monitor and facilitate delivery, that is robustly correlated with several indicators of health service delivery.

Overall correlations
To investigate whether characteristics of politicians and bureaucrats in Uganda's districts are robustly correlated with the quality of service delivery, we start by presenting simple

Assessing predictive power of all behavioral traits
While the previous section reports ndings from regressions that include each selection trait separately in equation (1) Following the literature, we proceed by reducing the "degrees of freedom" in model selection and instead relying on regularization techniques of machine learning 29 (Athey, district of Amudat. 27 The estimated e ect is 30% smaller than in the simple correlation, suggesting that part of the observed raw correlation is driven by other district characteristics that covary with politician integrity. 28 Results in regression form are presented in Table A6 29 We nonetheless report OLS estimates of di erent speci cations as robustness checks in Table 5. 2018). In particular, we use the LASSO procedure (Tbishirani, 1995)

Robustness
We start our robustness checks by estimating OLS regressions with Districts' DLT score as dependent variable and politicians' average integrity as explanatory variable in Table 5, and estimating models with increasing number of controls. Column (1) Table A8 32 Note that if individual traits are measured with more noise than Districts' overall characteristics, we should expect the coe cient on traits to be biased toward zero. tegrity barely changes, but standard errors become larger and the coe cient is no longer statistically signi cant (p-value = 0.11). As in the model chosen by LASSO, we nd that a higher number or radio headquartered in a district is a strong predictor of better health services, as is scores at the UCE national examination.
In column (4) we add the average integrity of bureaucrats as explanatory variable in order to assess if it has independent explanatory power. The coe cient on politicians' integrity remain unchanged, and that of bureaucrats' integrity is an order of magnitude smaller and not statistically di erent from zero. Given the large standard-errors of estimates, however, we cannot reject that these two coe cients are equal (p-value = 0.41). In a very demanding test, in column (5) we include all of bureaucrats' and politicians' traits simultaneously as explanatory variables, together with districts' characteristics. In this model, the coe cient on politicians' integrity is smaller (0.16) and not statistically di erent from zero. We should caution, however, that in this speci cation we are estimating 40 parameters with 75 observations. That is, the F test of a "horse race" between bureaucrat and politician integrity, is a weak test in our setting. We provide other tests below that reveal politicians' integrity, not that of bureaucrats', as a robust correlate of service delivery.
The available economic theory does not provide any predictions about whether bureaucrat or politician characteristics should matter more. We do not conclude from the results that bureaucrat traits don't matter, and indeed nd that bureaucrats' altruism does show-up as a signi cant correlate in several speci cations. Other bureaucrat traits might matter as well, but perhaps we are unable to discern it for a variety of reasons (we are underpowered; we have not measured the relevant traits well; local bureaucrats may be strategically appointed by national bureaucrats and politicians, making bureaucrat traits more endogenous to district characteristics which directly predict service delivery.) The one result that is revealed systematically by the data is that political quality (and speci cally, the characteristic of integrity) matters signi cantly for the productivity of bureaucratic organizations.
All results presented so far document how politicians' integrity correlates with overall health scores. The details of the health service delivery index construction are presented in Annex 9.3, but it's reasonable to assert that the weights attributed to each of the com-ponents, while linked to overall policy objectives, are somewhat arbitrary. This section decomposes the overall index in its individual indicators and evaluates the correlation between politicians' Integrity and each of the components. Figure 13 reports regression estimates of the coe cient on politicians' integrity, using the same speci cation as in column (2) of Table 5, but having each of the sub-components indicated as dependent variable. We standardize all components so that the coe cients are comparable across regressions.
The results reported in Figure 13 suggest that Politicians' integrity is robustly positively Higher politician' integrity is also positively correlated with measures of good administrative practice in the health sector, such as completeness and promptness of health facilities' reporting.
In a further robustness test on how much our nding depends on the measure of health outcomes, we replace the DLT score as dependent variable with the rst component of a Principal Component Analysis (PCA) of the 14 health indicators. The rst principal component used explains 22% of the variance in the sample 33 . We replicate the OLS speci cations from Table 5 in Table 6, but using the rst component as dependent variable. Results are broadly similar: estimates suggest that an increase of 1 s.d. in average politicians' integrity leads to an increase of 0.2-0.4 s.d. in the rst component of health outcomes. Note that here, using the rst component of health outcomes instead of the health score, the estimated coe cient on politicians' integrity remain statistically di erent from zero in all speci cations, even after including all of politicians' and bureaucrats' traits simultaneously. Furthermore, 33 It's worth noting that the PCA has negative factor loading in four our of our fourteen variables: TB success rate, Completeness of monthly reports, Completeness of facilities reports and Reports sent on time. This means that higher values in these variables, which should re ect higher performance, has a negative e ect on the rst component.
we reject in the speci cation in column (3) that the e ect of bureaucrats' integrity is the same as politicians' integrity (p-value < 0.1).
We also investigate whether the correlation between politicians' integrity and health outcomes holds when we measure integrity for di erent groups of politicians. Table 7 presents results where the explanatory variable is the average Integrity Index of di erent groups of politicians. Column (1) replicates the result that, once we control for several District characteristics, a 1 s.d. in integrity is associated with a 0.23 s.d. increase in the health score. Column (2) through (4) assess how this result changes if we calculate the average integrity for each politician type: local councillors, chairperson and political contenders.
Results are noisier and in no speci cation statistically di erent from zero. When we include all three integrity measures simultaneously, nonetheless, it is the average integrity of councillors that remains predictive: the coe cient is 0.27, very similar to estimates for the overall politician integrity, and the coe cients on both chairperson and candidates are much smaller and indistinguishable from zero 34 .
In Figure 11 we repeat a similar exercise as performed for overall integrity, assessing the correlation between politicians' traits and each of the subcomponents in the overall health index. The results suggest that it is indeed councillors' integrity which drive the correlation with good service delivery: not only councillors' integrity is signi cantly correlated with health indicators such as IPT2 and antenatal coverage, but point estimates are often larger than those obtained using chairperson's or contenders' integrity measures. These results, disaggregated by type of politician, are somewhat sensitive to including health spending and age of district as controls-the estimates of the correlation with councillor's traits becomes noisier for some of the individual health indicators, while the estimates of correlation with chaiperson's traits become more signi cant. 35 As a nal check of robustness, in Table 8 we examine whether other organizational 34 Note, nonetheless, that in this speci cation we lose 10 observations since we do not have chaiperson/candidate observations for all districts. 35 These results are not reported in the paper because including health spending reduces sample size by 3 observations, owing to missing values for 3 districts. The overall signi cance of average political integrity is unchanged. We discuss further in the next section that the sensitivity when including age of district is owing to signi cantly lower integrity of chairpersons in young districts, as erstwhile councillor positions are converted to chairperson positions with the carving-out of a new district from a councilor's seat. characteristics, that have been highlighted in the literature, matter in our data and whether our main results are robust to including them. First, in column (1)  We also examine whether ethnic heterogeneity among the districts' bureaucrats and politicians play a role in service delivery in columns (2) Finally, we consider whether the integrity of di erent types of bureaucrats-those with greater leadership powers, such as the Chief Administrative O cer (CAO) and the Resident District Commissioner (RDC), versus those within the technical health bureacracy, such as the District Health o cer-matters for service delivery, and whether this disaggregation by bureaucrat type changes the results for political integrity. We calculate the integrity index separately for the leaders of the bureaucratic organization -CAOs and RDCs -and for bureaucrats working in the health sector, and include these in the speci cation reported in column (5). The coe cient on bureaucratic leaders' integrity is very small in magni-36 It includes six questions on whether bureaucrats believe that senior management recruit and promote the best employees. tude, while that for health bureaucrats is somewhat larger but noisily estimated. More importantly, we see no di erence in the results on politicians' integrity.

Correlates of integrity
The previous section documented a robust correlation between politicians' integrity, in particular that of local councillors', and health outcomes. While we are not in a position to identify an exogenous source of variation in political integrity, we would like to explore where variation in integrity might be coming from by exploring its covariates in the data.
Our question in this section is therefore: which variables predict high integrity among politicians?
We start this analysis by presenting disaggregated data on behavioral traits by type of politician; do Councillors and elected Chairs, or contenders vs. elected Chairs, di er systematically from one another? Table 9 presents average characteristics of Chairpersons, non-elected contenders for Chairperson and Councillors. The rst overall nding is that Councillors underperform elected Chairpersons and contenders for that position: they are much less educated, perform worse in the cognitive test and score lower in Integrity and Non-Cognitive measures. Among those who ran for Chairperson, di erences are less systematic between winners and losers: winners seem to perform slightly worse in Integrity and Non-Cognitive measures, but estimates are noisy. The one signi cant and meaningful di erence is in the asset index, suggesting winners come from wealthier households than those defeated.
We then proceed to evaluate whether politicians' integrity is consistently correlated with Districts' socio-economic and political characteristics and individuals' observable traits. Column (1) of Table 10 presents estimates for all politicians, while columns (2)-(4) restrict the sample to Chairpersons, Contenders to Chairperson and Councillors, respectively. Two results are worth noting. One, electoral competition is associated with higher political integrity among winning candidates; and two, younger districts tend to have worse integrity among chairpersons. Lower competitiveness, as measured by the average Her ndahl index for councillors, is correlated with worse integrity in the pooled sample, as well as separately for elected chairpersons and councilors (signi cant at the 10 percent level). Chairperson integrity is lower in younger districts-those that are carved out of sub-district units which used to be the seats of councillors. This result is consistent with the general nding that councillors tend to have worse integrity than chairpersons, as chairpersons in younger districts are more likely to come from the pool of councillors.
Theoretical models of candidate entry into political markets (Caselli and Morelli, 2004;Mattozzi and Merlo, 2008;Leon, 2013) may be extended to account for the empirical pattern we observe-that politicians in lower-level positions have worse integrity. Politicians in lower-level positions of councillors face greater challenges in signaling their type to voters, because their actions and impact on service delivery outcomes are harder to discern, since they have fewer formal powers over service delivery compared to higher-level actors. For example, lower-level politicians can claim that service delivery failures are all due to lack of funding being provided by senior district and national-level leaders. Therefore, following the logic of the theoretical literature, "bad" types may be more likely to enter as candidates, and win, in lower-level positions compared to higher levels. However, other strands of literature would suggest that voters have more information about more local politicians, perhaps because they are closer to their social networks (Munshi and Rosenzweig, 2008;Casey, 2015). Our nding of lower integrity among more local politicians is therefore a puzzle that cannot be easily explained by arguments available in the literature.
In other results, politicians from more urban Districts score worse on the integrity index. Older politicians and those with a post-graduate degree score higher, on the other hand. These results, however, are not very robust when we restrict the sample to di erent kind of politicians: standard errors of estimates become very large and coe cients often change signs. Furthermore, estimates for region xed-e ects are not signi cant, suggesting that average integrity of politicians does not vary substantially after accounting for other observables. Therefore, the evidence suggests that, as far as measured, politicians' integrity is not systematically correlated with Districts' socio-economic characteristics (that is, characteristics other than electoral competition).
We also examine the correlation between integrity and other behavioral traits. In documenting bivariate correlations between integrity and a wealth of personal characteristics, we follow Cantone et al. (2016) and present p-values adjusted for multiple hypotheses testing. Table 11  Having shown that several of the behavioral traits are systematically correlated with higher integrity among politicians, we proceed to evaluate whether these individual correlations hold when we keep xed other characteristics. Table 12  The pattern of correlation between integrity and Conscientiousness, and between integrity and one particular component measure of the Public Service Motivation index-Commitment to Public Interest-suggests that the in uence on bureaucratic productivity of political selection is occurring through greater intrinsic motivation to perform assigned duties (Conscientiousness), as well as, possibly, greater pro-social preferences (Commitment to Public Interest) which motivate agents to deliver public goods such as public health.
Future work on political and bureaucratic selection in local bureaucracies could focus on better measuring pro-social preferences, and attempt to disentangle the roles of di erent sources of non-pecuniary motivation, whether arising from integrity or conscientiousness to do one's job as tasked, or, preferences for public goods. The currently available measure of Public Service Motivation is drawn from a literature focused on American bureaucracies and the political context of the United States (Perry, 1996) and needs greater testing and validation in poor country contexts. As mentioned before, we regard the results we present here as just a start, albeit an important rst step, towards closing the empirical gap between theory and empirics on the role of non-pecuniary motivation in shaping productivity of bureaucracies.

Discussion of mechanisms
The interpretation we o er of the robust correlation between political integrity and health service delivery is that in districts where higher integrity politicians are endogenously selected, these politicians are more motivated to perform their multiple roles within the complex organization of district bureaucracies to improve public health, including sensitizing citizens to the value of health services, and monitoring and supporting health workers at the frontlines to deliver. As discussed in section 2, government institutions in Uganda explicitly provide a role to locally elected politicians to monitor and facilitate service delivery, working as liaisons between citizens and the technocrats in charge of delivering nationally-mandated programs. In particular, primary health and education services in Uganda have been examined in prior research as areas where local politicians exercise inuence (Akin et al, 2005;Reinikka and Svensson, 2004;Brosio, 2000). This prior research has, in fact, raised concerns about local political accountability for providing public goods and managing scal resources. For example, Akin et al (2005) argue that district governments are more likely (than the center) to spend on private goods rather than public goods. Brosio (2000) refers to opinion surveys of local councilors in Uganda to suggest that these councilors are more concerned about paying themselves allowances rather than ensuring that health facilities are well equipped with drugs and implements. In a celebrated study of "leakage" in public spending, Reinikka and Svensson (2004) suggest that local politicians diverted funds from a national transfer of capitation grants intended for schools, and instead used these funds for their political campaigns. In such an environment, higher integrity politicians can be reasonably expected to di er in their motivation and therefore in the degree to which they positively in uence service delivery. However, there are competing, alternate explanations that we consider below.
Perhaps the most important alternate explanation is that underlying (unobserved) conditions of household demand for health, or social norms that facilitate health service delivery, may be simultaneously shaping the selection of politicians and service delivery performance, with no direct connection between the two. Other explanations centered around omitted variables, that local economic and political conditions are driving both political selection and health services, are somewhat attenuated by our ability to control for local area poverty estimates, education, and electoral competition (as reported in the results section). However, citizen demand and social norms are harder to account for. We argue that two variables-locally headquartered radio stations, and vote share of the national political leader-are likely to go a long way towards mopping-up the direct in uence of citizen demand and social norms. These are, in fact, the only two other variables, other than political integrity, that are robustly signi cant in explaining variation in health service delivery.
However, the signi cance of political integrity survives the inclusion of these variables. This suggests that political selection plays an additional and independent role in shaping service delivery, through the behavior of politicians. Nevertheless, the coe cient on political integrity that we are able to estimate can be biased because of omitted variables, such that our results cannot be extended to predicting the size of any service delivery impact of random assignment of politicians of higher integrity.
What about reverse causality? This is the other problem to consider when exploring how to interpret a robust empirical correlation in the data. Would places with better health services be more likely to select better politicians? There is nothing in the voluminous literature on Uganda that would support such an interpretation. One reason to doubt it is that the health services being measured are related to maternal and child mortality, rather than to adult human capital, such as through education, which may shape political behavior such as deciding whether to vote for corrupt or clean candidates. We use data on districts' average students' score at the UCE, a nationally administered exam students take at the end of lower secondary school, as the best available proxy for education in the population.
There is indeed a positive correlation between better education outcomes and better health service delivery. However, including or excluding the education variable does not change the coe cient on political integrity.
Another way to think about reverse causality is consistent with our preferred interpretation of mechanisms (that politicians of higher quality take better actions to improve service delivery): if experience with better health services gives citizens greater trust or con dence in local institutions, and thus makes them more likely to select clean candidates, part of the reasons why they would do so is because of the expectation (that should be ful lled if this were the equilibrium) that clean candidates will be able to do a better job of ensuring service delivery. There is a speci c role district councilors are supposed to play to monitor health facilities. If active citizens choose better quality politicians, by that very logic, we may reasonably expect them to choose politicians who can do a better job monitoring the health facilities. Hence, the proximate reason behind the correlation of political integrity and service delivery would, logically, be more active politicians.
In ongoing work in a seperate research project in the district of Hoima, we discover qualitative evidence of the role of local politicians (Ferraz, Habyarimana, and Khemani, ongoing). We held meetings with district bureaucrats and politicians and asked them to deliberate upon the most important problems of their district and what they could do to improve the situation. Health services were identi ed by the local politicians in Hoima as the most important problem, and the role of local politicians was highlighted to empower technical o cials to deliver it. Interestingly, our data shows that Hoima's district councilors have signi cantly higher integrity than councilors in other districts, although they are undistinguished on other selection traits.
In the course of the deliberations in Hoima, the District Health O cer is recorded as saying: "We as technical people, we need support from di erent leaders", to which the directly elected district chairperson responds as follows: "And as leader, I want to promise total political support to my technical team, where it is necessary. . . political support, we shall not interfere with what you are doing in line with health. So we should also empower our health workers, down on ground that you are in charge, feel yourself that you are in charge and when you're commanding for example the health in-charge your sta , let them feel that you are the boss of that health facility and they are supposed to do what you have directed them to do. Be empowered that you are the boss and you want something to happen as we have agreed here. That empowerment is what am talking about". In another part of the discussion, the same politician talks about the importance of lower-level politicians-the councilorsmonitoring health facilities: "Our parish chiefs must make sure that they are monitoring our facilities and write reports, the parish chief can report to our sub county chief, then the sub county chief writes a report to the DHO, then the DHO writes a report to the CAO. So that now the CAO can present such ideas for action and adjustments where necessary. So that is what we are now going to plan to put into place. . . to enforce because it is in the existence...". This exchange suggests that local politicians can play both negative and positive roles in the delivery of services-interfering to obstruct versus monitoring to facilitate better services.
The integrity of local politicians can be expected to matter in whether they obstruct or support the technical o cers in delivering services within the complex organization of local government. For example, if local politicians are colluding with health service providers to extract rents from the health facility, they may obstruct disciplinary action by district management. In another study, in Pakistan, Callen et al (2014) nd that doctors with connections to political leaders are more likely to be absent from public health clinics, and the public o cials who manage these doctors are more likely to report political interference when trying to apply sanctions. In a study in Uganda, in the district of Arua (which is in our sample, and has signi cantly lower average political integrity compared to Hoima), Grossman, Platas and Rodden (2018), report qualitative evidence that the Medical O cer In-charge of a health center had been absent for several months, and that it took a community-wide meeting facilitated by external actors to get this issue to be reported to the district leaders and prompt them to take disciplinary action. In such a setting, the in-tegrity of local politicians who are supposed to monitor and make reports to management, can play a direct role in whether the politicians take actions that improve or discourage service delivery. Grossman, Platas and Rodden (2018) also provide qualitative evidence that district bureaucratic leaders rely on local councilors to resolve service delivery problems raised by citizens, since these councilors are located at the sub-district level where service delivery happens, and are equipped to mediate between citizens and frontline service providers. The results of Groassman et al's (2018) evaluation of a citizen engagement intervention in Arua shows that these interventions targeted at increasing citizen activism did not impact health service delivery. This evidence supports our interpretation that citizen activism alone cannot explain the correlation between political quality and service delivery; actions by local politicians are likely to be the proximate route through which political quality matters.

Conclusion
This paper contributes to understanding local state capacity to deliver public goods. In a recent review, Dal Bo and Finan (2016) list as an open question the e ect of decentralization on the delivery of public goods, despite more than two decades of policy e ort towards establishing local governments in poor countries. The research agenda has shifted away from comparing centralized versus decentralized provision of public goods towards examining principal-agent relationships within myriad complex organizations of government.
Key ndings from available research are in the nature of con rming basic economic principles such as: pay for performance tends to strengthen incentives and improve service delivery (Banerjee et al, 2008); higher wages can improve the selection of state personnel (Dal Bo et al, 2013); and monitoring can reduce corruption (Olken, 2007).
The open question which is particularly relevant for policy-makers in Africa and other developing regions is how to structure principal-agent relationships in frontline jurisdictions of government, that is, the local governments which implement national policies and deliver spending programs. Local bureaucracies are tasked with multiple complex jobs, such as managing cadres of health workers and teachers to deliver basic health and education services on which poor households rely, and implementing infrastructure programs for local agriculture and economic growth. Finan et al (2015) conclude their review of the personnel economics of the state with the observation that when government bureaucracies have to perform multiple complex tasks, questions of how to structure the principal-agent relationships within these bureaucracies receive few answers from existing research. This paper is beginning to ll the gap.
We provide evidence consistent with economic theory of complex principal-agent relationships in which non-pecuniary motivation and selection of di erent types of agents in uence organizational productivity. We gather unique survey data to measure selection traits-integrity, altruism, personality, public service motivation, risk aversion, cognitive ability-of bureaucrats and politicians who make-up the complex organization of local government to provide the rst evidence of whether (which) selection traits are correlated with organizational productivity. We provide robust evidence that it is political integrity in particular that matters for service delivery, to the exclusion of other traits. This evidence is particularly relevant for understanding local state capacity for policy implementation in poor countries. This is because budget and human resource constraints in poor countries can limit the extent to which highly trained professionals can be attracted to the public sector, or be given su cient incentives to make them accountable for their performance.
On the other hand, politics is ourishing in poor countries-with widespread participation of citizens as voters and as contenders for political leadership at local levels (World Bank, 2016a)-whether for ill or good. The evidence in this paper shows that local politicians can exert signi cant in uence over the functioning of bureaucracies at the frontlines. Policy e orts to build capacity of the local state would do well to take seriously the role of political selection, going beyond incentives, when crafting the principal-agent relationships (who is tasked with doing what) in local bureaucracies.

Bureaucrats Selection Traits
Note: This graph reports point estimates and 95 percent CI for the coe cient on "Politician" dummy variable, in regressions with selection traits as dependent variable and district xed-e ects and individual characteristics (gender, age education, household size and father education) as controls. Each reported coe cient refers to a separate regression.     Note: These graphs report the point estimate and 95 percent CI for the coe cient on selection traits in regressions using District level DLT Score as dependent variable. Each coe cient refers to a separate regression with non-reported controls including regional dummies, log total population, poverty rate, share of urban population, number of radio stations' HQs, vote share of the presidential candidate of the NRM, her ndahl index for councillors' votes and average students' score on UCE. CI are constructed using robust standard errors.  Note: This graph reports the point estimate and 95 percent CI for the coe cient on the variable in the axis, in a bivariate regression using District level health outcomes as dependent variables. CI is constructed using robust standard errors. The variables used here are those previously selected in a LASSO model. All variables are standardized to have unit standard deviation.  Note: This graph reports the point estimate and 95 percent CI for the coe cient on Average Politician Integrity in regressions using District level DLT Score as dependent variable. Non-reported controls include regional dummies, log total population, poverty rate, share of urban population, number of radio stations' HQs, vote share of the presidential candidate of the NRM, her ndahl index for councillors' votes and average students' score on UCE. CI are constructed using robust standard errors. CI is constructed using robust standard errors.  Note: This table reports regressions using District level health outcomes as dependent variables. Nonreported controls include regional dummies, log total population, poverty rate, share of urban population, number of radio stations' HQs, vote share of the presidential candidate of the NRM, her ndahl index for councillors' votes and average students' score on UCE. CI are constructed using robust standard errors. * p < 0.1, ** p < 0.05, *** p < 0.01. Standard errors in parentheses.  Figure 11: Health outcomes vs. Integrity by politician's type Note: These graphs report point estimates and 95 percent CI for the coe cient on Average Politician Integrity in regressions using District level DLT Score as dependent variable. Non-reported controls include regional dummies, log total population, poverty rate, share of urban population, number of radio stations' HQs, vote share of the presidential candidate of the NRM, her ndahl index for councillors' votes and average students' score on UCE. CI are constructed using robust standard errors. CI is constructed using robust standard errors.

Survey Data -Construction of Individual Selection Variables
Public Sector Motivation (PSM) -In order to measure bureaucrats' and politicians' motivation to serve in the public sector, we use Perry's Public Sector Motivation Index (Perry, 1996). Annex I presents the version used in the survey in Uganda: it is composed of 36 questions measuring motivation in six di erent dimensions -Attraction to Policy Making, Commitment to Public Interest, Social Justice, Civic Duty, Compassion and Self-Sacri ce.
As discussed by Perry (1996) to 5 (Strongly Agree). We recode the variables such that higher values are disagreement with unethical behavior, then proceed to take a simple average of the eight questions and normalize it to obtain the Integrity Index.
Risk Aversion -Risk preference is one of the fundamental parameters to understand individuals' dynamic decisions (savings, consumption, investments in human capital, etc. . . ).
We measure risk preferences by using a common risk game: individuals are given the possibility of choosing, in a hypothetical game, between a certain gain of UGX 1 million and a gamble with 50/50 probability with expected value higher than UGX 1 million. If the individual chooses the certain prize the game ends; if they choose the gamble, they are given another bet between the same certain prize and a lottery with lower expected value.
The game can last up to 4 rounds, giving us ve levels of risk-aversion (See Annex IV for details). In order to screen if individuals understand the game, an initial question is asked giving the interviewee the decision between a certain prize of UGX 1 million and a 50/50 bet between UGX 1 million and UGX 2 million. Choosing the previous is clearly dominated by the latter, so individuals who choose that even after given a chance to reconsider are marked as choosing a dominated risk option. We follow Dal Bo, Finan and Rossi (2013) in using that choice as another measure of cognitive ability.
Cognitive Ability -The survey uses a digit span test to proxy for cognitive / memory ability. We use a similar module to Weaver (2016), I which the interviewer reads out a string of digits, starting with three, and the respondent has two chances to repeat them right. If they can't, the test ends; if they do remember, they are given a longer string; the game goes on until the longest string with 12 digits. We code how many digits the individual could remember, given two chances. As discussed above, we also use mistakes in choosing the dominated option in the risk game as a measure of cognitive ability.
Altruism -We measure altruism by the percentage of money shared in a hypothetical game in which the individual is given UGX 50,000 and has to decide how much to keep for themselves and how much to share with an anonymous person who received no money.

Reliability of selection traits
Here we brie y discuss the reliability of our three composite indices of selection traits: , where K is the number of items;c is the average variance of items andv is the average covariance between all components. Cronbach's alpha ranges from zero to one, and by construction if tends to one as the number of items increase. 0.6, and average inter-item covariance from 0.5 to 0.12. The overall PSM index, however, shows higher Cronbach's alpha of 0.76.
The Integrity index, comprised of eight items, has an alpha of 0.61 and average interitem covariance of 0.14.
Each of the Big Five components show rather low alphas, between 0.2 and 0.4. These results, however, are consistent with those found by Pierre et al. (2014) using the same instrument in di erent low-income countries. It should be noted that we use a simpli ed Big Five Instrument, with only 15 questions instead of the usual 44-question instrument.
The overall Big Five Index presents moderate reliability (alpha = 0.57). Finally, the overall Non-Cognitive Index shows higher reliability with alpha = 0.68.  The Her ndahl Index is most commonly used as measure of rm competition in a given I am now going to ask you a series of questions based on the following scenario: Imagine you can choose between two bags. Once you have chosen one of the bags, you will put your hand inside the bag and without looking you will pick a ball which will show the amount of money you have won.

Chooses dominated Risk Option: if individual chooses Bag No.1 in both questions
below, it's agged as choosing dominated risk option.
1. Now supposing Bag 1 has one ball that is worth UGX 1 million and Bag 2 has two balls: one is worth UGX 1 million and the other ball is worth UGX 2 million. Which one of these two bags do you choose; Bag 1 or Bag 2?
2. Are you sure? You are going to pick only one ball from the bag you chose. Now if Bag 1 has two balls each worth UGX 1 million. If you choose Bag 1 you will win UGX 1 million.
If you choose Bag 2 you will win at least UGX 1 million or UGX 2 million depending on your luck. Which one of these two bags do you choose; Bag 1 or Bag 2?

Risk-Aversion
The questions below are asked in sequence, whenever Bag 1 is chosen the game end.
Choices are coded as follow: • If Bag 1 is chosen in rst question: Very High Risk Aversion; • If Bag 1 is chosen in second question: High Risk Aversion;     Note: This tables reports results from regressions having District DLT score as dependent variable and each of politicians' average selection traits plus a series of controls as explanatory variables. Non-reported controls include regional dummies, poverty rate, log of total population, share of urban population, UCE scores in 2011, vote share for NRM in 2011, Her ndahl index for councillors' votes in 2011 and number of radio HQ in District. The coe cients on this table are presented in Figure 8 . Note: This tables reports results from regressions having District DLT score as dependent variable and each of bureaucrats' average selection traits plus a series of controls as explanatory variables. Nonreported controls include regional dummies, poverty rate, log of total population, share of urban population, UCE scores in 2011, vote share for NRM in 2011, Her ndahl index for councillors' votes in 2011 and number of radio HQ in District. The coe cients on this table are presented in Figure 8. Note: This table presents regressions using districts' health outcomes as dependent variable and individual traits and districts' characteristics chosen by the LASSO method as explanatory variables. Columns (1) -(9) present the results from bivariate regressions (equivalent to Figure 9a), while column (10) presents the multivariate model (equivalent to Figure 9b)