Policy Research Working Paper                8644




               Hierarchy and Information
                             Daniel Rogger
                             Ravi Somani




Development Economics
Development Research Group
November 2018
Policy Research Working Paper 8644


  Abstract
 The information that public officials use to make decisions                        accuracy. The errors of public officials are large, with 49
 determines the distribution of public resources and the effi-                      percent of officials making errors that are at least 50 percent
 cacy of public policy. This paper develops a measurement                           of objective benchmark data. The provision of briefings
 framework for assessing the accuracy of a set of fundamen-                         reduces these errors by a quarter of a standard deviation, but
 tal bureaucratic beliefs and provides experimental evidence                        in line with theoretical predictions, organizational incen-
 on the possibility of ‘evidence briefings’ improving that                          tives mediate their effectiveness.




 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 drogger@worldbank.org.




         The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development
         issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the
         names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those
         of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and
         its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent.


                                                       Produced by the Research Support Team
                                 Hierarchy and Information

                                        Daniel Rogger and Ravi Somani∗




                           The most recent version of this paper can be found here.




          Keywords: information, bureaucrats, public administration, decentralization
          JEL Classiﬁcation: D73, D82, H00, H11, H70, H83, O20




  ∗
     Rogger: World Bank Research Department, 1818 H St NW, DC 20433, USA, and Institute for Fiscal Studies (IFS);
drogger@worldbank.org. Somani: Department of Economics, University College London; ravi.somani.09@ucl.ac.uk. We
thank the Ministry of Public Service and Human Resource Development of the Federal Government of Ethiopia for their
leadership and support throughout design, implementation and data collection. We are very grateful to Elsa Araya, Verena
Fritz and Kerenssa Kay of the World Bank’s Governance Global Practice for collaboration on implementation of the survey
that underlines this work. We gratefully acknowledge ﬁnancial support from the Centre for the Study of African Economies
[grant number CUR00090], the UK’s Department for International Development, the International Growth Centre [grant
number 32407], and the World Bank’s Development Impact Evaluation i2i fund, Bureaucracy Lab and Governance Global
Practice. We are grateful to Pierre Bachas, Guadalupe Bedoya, Antonio Cabrales, Aidan Coville, Stefan Dercon, Eliana La
Ferrara, Arianna Legovini, David McKenzie, Macros Vera Hernandez, and seminar participants at Bocconi, George Washing-
ton University, Paris School of Economics, Kent, Namur, Oxford and the UK’s Department for International Development
for valuable comments. All errors remain our own.
1        Introduction

The information that public oﬃcials use to make decisions determines the distribution of public resources
and the eﬃcacy of public policy (Callen et al., 2018; Dal Bó et al., 2018). Errors in that information
can reduce social welfare as the marginal eﬃciency of public investments falls. Correspondingly, policy-
related research frequently attempts to improve the information set of public oﬃcials and thereby raise
the eﬃciency of policy.1
        Despite the importance of bureaucratic beliefs for citizen welfare, there is an absence of empirical
evidence on the accuracy of information used in public sector decision-making. This paper develops a
measurement framework for assessing the accuracy of a set of fundamental bureaucratic beliefs and provides
descriptive evidence on their determinants. We then provide experimental evidence on the possibility of
improving the accuracy of those beliefs through the provision of evidence brieﬁngs, akin to those used by
actors inside and outside the public service to inform public oﬃcials.
        More precisely, we develop a survey that extracts the beliefs of public oﬃcials about the key variables
on which their work depends, such as the size of the population they are serving or sector-speciﬁc variables
such as the number of primary-age children who are in school. We compare the responses of the surveyed
oﬃcials to objective benchmark data from administrative and survey data to calculate the scale of errors
that these oﬃcials make. We thus provide among the ﬁrst large-scale descriptive evidence on the nature
and correlates of information in the public sector.
        We combine these measures with information about the structure of the hierarchy in which public
oﬃcials work and the management practices they operate under to show that our descriptive results
closely mirror a series of core theoretical predictions regarding the distribution of information in the
public sector. Information acquisition in bureaucratic hierarchy is theoretically distinctive to models of
learning by autonomous agents. Unlike markets, bureaucracies do not have organic information creation
mechanisms, such that individual oﬃcials must undertake costly actions to acquire and absorb information.
As bureaucratic decision-making is based on team production and decision-making, oﬃcials face the threat
of being over-ruled by their superiors and an incentive to free ride on the information acquisition of other
bureaucrats (Argenziano et al., 2016; Di Pei, 2015; Stiglitz, 2017). Our empirical results closely match the
predictions of this theoretical literature.
    1
     This is an explicitly stated aim of many important economics research centres such as: the US-focused Bipartisan
Policy Center’s Evidence-Based Policymaking Initiative, which provides “advice and expertise on implementation options
and strategies ... as Congress develops legislation and as the Executive Branch devises regulations, policies, and standards
to improve the generation of high level, quality evidence.”; the Centre for Economic Policy Research (CEPR) that “initiates,
funds and coordinates research activities and communicates the results quickly and eﬀectively to decision makers around the
world.”; the Abdul Latif Jameel Poverty Action Lab (J-PAL) which is “a global research center working to reduce poverty by
ensuring that policy is informed by scientiﬁc evidence.”; and the World Bank, that “generates high-quality and operationally
relevant data and research to transform development policy”.



                                                             2
   We enumerate our survey to a representative sample of oﬃcials across the Government of Ethiopia.
Our survey spans 1,831 oﬃcers working in 382 organizations spanning all three tiers of government, and
the variables on which we collect data are fundamental quantities of their work. Given the novelty of our
data and the scarcity of empirical evidence on this subject, we provide a substantial amount of descriptive
evidence on the nature and scale of errors in claims made by public oﬃcials of Ethiopia’s government.
   A minority of the public oﬃcials we study make relatively accurate claims about their constituents. Of
oﬃcials’ assessments of the population they serve, 22% are within 20% of the census-deﬁned population.
However, a large proportion of bureaucrats make economically-meaningful mistakes about the basic con-
ditions of local jurisdictions. 50% of oﬃcials claim that their district’s population is 50% bigger or smaller
than it is. The scale of errors is similar across the distribution of district populations, implying that public
oﬃcials are indeed adjusting their estimates towards the underlying state of nature, but doing so with a
lot of noise. Across the socio-economic variables we study, 49% of public oﬃcials make errors that are
at least 50% of the underlying benchmark data. Restricting our analysis to that set of variables that are
the primary responsibility of oﬃcials, such as those questions asked to education oﬃcials about education,
62% of oﬃcials make errors that are at least 50% of the underlying benchmark data. Such large errors are
consistent with studies of agents from other settings, such as farmers not optimizing important aspects
of the technology they use (Hanna et al., 2014), Indian manufacturing ﬁrms having limited information
on quality defects, machine downtime, and inventory (Bloom et al., 2012), and professional forecasters
making large but predictable mistakes (Ehrbeck and Waldmann, 1996; Andrade and Le Bihan, 2013).
   These results matter because public oﬃcials state that their tacit knowledge is a key source of infor-
mation in public policy decision-making. As Duﬂo (2017) argues, policymakers “tend to decide on [policy
design details] based on hunches, without much regard for evidence.” Our survey asks where oﬃcials
source their knowledge to make public policy decisions. Consistent with other studies (World Bank, 2012;
Masaki et al., 2017) we ﬁnd that bureaucrats rely dominantly on their own, and their colleagues, tacit
knowledge. Oﬃcials’ personal beliefs over the characteristics of the citizens they serve therefore matter for
resource allocation decisions and the eﬀectiveness of implementing public policy. The errors we document
imply substantial skewness in how public resources are directed. In a simple scenario in which we model
the public oﬃcials we study equitably distributing resources across the constituencies they serve, districts
at the 75th percentile of errors made by oﬃcials receive 30% more funds than their actual population
warrants. Compounded by further errors on the nature of citizen need, such as indicators of health status,
our data imply that public oﬃcials are poorly targeting public resources due to a lack of information.
   To our knowledge our measurement framework is the ﬁrst direct measure of individual-level information
in related literatures, and we apply it across a large-scale public sector hierarchy. Fisher et al. (2018) record



                                                       3
‘knowledge-seeking’ activities (proxied by downloads from and visits to a particular website). The ﬁnance
literature measures the (‘soft’) information held by loan oﬃcers at lending institutions. The measures of
soft information are either elicited from survey questions that ask directly the extent to which employees
feel that they know borrower characteristics (Uchida et al, 2012; Hattori et al, 2015), the predictive power
of internal risk ratings relative to ‘hard’ measurable characteristics of ﬁrms (Agarwal and Hauswal, 2009,
2010; Brown et al, 2012; Qian et al 2015; Agarwal and Ben-David, 2018, Dal Bo et al, 2018), or measures
of the eﬀort exerted by relationship managers to learn about their clients (Liberti, 2018).
       We show descriptively that providing oﬃcials with de facto control over their tasks improves the preci-
sion of their claims about constituents by a third of a standard deviation. This accords with corresponding
theoretical predictions (Aghion and Tirole, 1997; Garicano, 2000; Dessein, 2002) and uses the fact that
the Ethiopian government is organized around multiple tiers of government oﬃces to proxy the delegation
of de facto authority across tiers. By limiting our comparisons to claims about the same indicator, at
the same percentile of its distribution, for the same district, and controlling for individual factors such
as education and experience, we are able to provide the most precise descriptive evidence to date on the
relationship between delegation and information (Acemoglu et al., 2007; Aghion et al., 2017).2
       Similarly, the value assigned to the use of information within an organization mediates the impacts of
the delegation of authority (Gant et al., 2002; Mookherjee, 2006). In those organizations that have weak
incentives for information acquisition, the delegation of authority may have limited impact since there is
little impetus for actors to use it. Using measures of management practices as a proxy for organizational
incentives, we ﬁnd that the delegation of de facto authority has a positive eﬀect on the acquisition of
information only when the appropriate management practices are in place. District oﬃcials working in
those organizations with the weakest management practices related to monitoring are no more informed
than their central colleagues. This implies that local oﬃcials are not always better informed, as often
assumed in the literature on optimal delegation and decentralization (Oates, 1999; Bardhan, 2002), but
are better informed only if there are appropriate incentives in place for information acquisition.
       Building on this descriptive work, we experimentally investigate whether the accuracy of bureaucratic
beliefs can be improved by evidence brieﬁngs that are a frequent means by which government or academic
research is communicated to public oﬃcials. Such brieﬁngs have long been used within public service but
are increasingly a focus of research entities such as the Centre for Economic Policy Research, JPAL/IPA,
and the World Bank Group. For example, the number of evidence brieﬁngs produced per year by the
Campbell Collaboration, a focal coordinator in the “use of systematic reviews and other evidence synthesis
   2
    The variation in the precision of public oﬃcials’ estimates does not seem to be mediated by basic personal characteristics,
such as an oﬃcial’s education or experience, or by their proximity to the local environment we are questioning them on.
Demographic variables and proxies of the remoteness of the oﬃcials from their constituents do not explain a substantial
proportion of the variation in errors.


                                                              4
for evidence-based policy”, has increased steadily over the past decade (Campbell Collaboration, 2019).
    There is surprisingly little evidence on whether such brieﬁngs do in fact inform public oﬃcials. Callen
et al (2018) document changes in frontline outcomes from the installation of a new information technology
in Pakistan. Elliott and Popay (2000) present qualitative evidence from the UK that organizational factors
appear to signiﬁcantly mediate the impact of research in healthcare organizations. This study is the ﬁrst
to test the direct eﬀect of evidence brieﬁngs on bureaucratic beliefs.
    We provide an information package that exactly mimics circulars (internal communications) of the
Ethiopian government to a random sub-set of the oﬃcials that we study. The information package contains
all the information we test for in our survey. Lowering the marginal cost of information acquisition
dramatically and symmetrically in this way should yield superior information overall (Aghion and Tirole,
1997).3
    We ﬁnd an average treatment eﬀect equivalent in magnitude to the descriptive impact of de facto
delegation. The errors made by treated oﬃcials are 0.28 of a standard deviation lower than control oﬃcials.
Our results are consistent with evidence from a range of settings that reducing the cost of becoming
informed increases information acquisition. Hastings and Weinstein (2008) ﬁnd that providing a simple
information packet to parents regarding local school test results increases the proportion of parents that
factor this information into their school choice. Jessoe and Rapson (2014) ﬁnd that electricity consumers
become more responsive to price changes when presented with high-frequency information about their
residential electricity usage through a simple in-home display. Banerjee et al. (2018) show that mailing
Indonesian households information on their eligibility for rice subsidies resulted in a signiﬁcant increase
in the amount of subsidy that they received, implying that they absorbed and acted on the information.4
Our experiment extends these results to public policy making, in which free-riding concerns bring in to
question whether even large decreases in the cost of acquiring information have impacts on its equilibrium
level (Aghion and Tirole, 1997). We provide evidence that it does.
    However, the treatment eﬀect arises totally from organizations with below-average management prac-
tices, reinforcing our descriptive ﬁndings that organizational incentives are key mediating factors in how
new information is received and absorbed. While external interventions can reduce the scale of errors made
by public oﬃcials, suitable organizational incentives ensure self-acquisition of that information. This result
is of signiﬁcance to researchers who aim to update the beliefs of public oﬃcials. Organizational incentives
    3
      This ﬁnding is echoed in Gabaix et al. (2006), where information acquisition is a decreasing function of both the cost
of acquisition and its complexity, and Garicano (2000), in which cheap information is widely shared within an organization.
Myatt and Wallace (2012) study how much costly attention players in a coordination game should pay to signals that vary
in accuracy (how precisely it identiﬁes the state) and clarity (how easy it is to understand), ﬁnding that the clarity of signals
is of primary importance even if they have poor underlying accuracy.
    4
      In the Banerjee et al study, the more information that was provided on the mailing, the larger the treatment eﬀect,
implying that it was the information, rather than simply the presence of a reminder that was driving the results.



                                                               5
to acquire information determine the eﬃcacy of interventions and technologies aimed at improving the
information of public-sector agents. Our experimental ﬁndings highlight that when the public sector is
designed appropriately, the rationale for compensating interventions - which are often ad hoc and focused
on short term goals - is weakened.5 The paper thus relates to the nascent literature on how the de-
sign of bureaucratic organizations impacts on state capabilities (reviewed by Finan et al., 2017), and on
the importance of these capabilities for development outcomes (Besley and Persson, 2011; Acemoglu and
Robinson, 2012; Pepinksy et al., 2017).6
    In line with much of the recent literature on bureaucrats (such as Dal Bó et al. (2013) and Bertrand
et al. (2016)), our focus is to understand the determinants of the characteristics of bureaucracy, rather
than its wider eﬀects. However, we ﬁnd some evidence that superior information improves promotion
prospects of bureaucrats and, in the form of stating knowledge of the indicators we focus on, improves
service delivery outcomes. This is consistent with descriptive evidence from our survey that information
held by civil servants on the conditions in the constituency they serve feeds into allocation and operational
decisions. Oﬃcials in our survey state that on 74% of projects, “information regarding the state of oper-
ations and services in the jurisdiction makes it easier for me to know where to allocate my eﬀort/time”.
Similarly, 82% of oﬃcials stated that jurisdictional information matters for a majority of the operations
(projects/programs) that they work on. It is also consistent with Callen et al. (2018) who document
increases in doctor attendance from a system that highlights poorly performing districts in Pakistan to
Executive District Oﬃcers and Dal Bó et al. (2018) who provide evidence that the superior local in-
formation of middle managers improves the targeting of monitoring interventions aimed at agricultural
extension workers in Paraguay.
    At the same time, the scale of errors made by oﬃcials does not systematically predict the service deliv-
ery indicators we have information on. This ﬁnding is consistent with Dhaliwal and Hanna (2017) in which
detailed information on the absence rate of doctors provided to local oﬃcials in India has limited eﬀects
on service delivery. Together, these results highlight the need for more research on how the information
of public oﬃcials and its equilibrium distribution in government relate to the delivery of public goods.
    The rest of this paper proceeds as follows. Section 2 details the approach we took to measuring
information in the public sector. Section 3 presents detailed descriptives of the determinants of information
    5
      The results echo those of Olken (2007) in which top-down government monitoring reduces corruption substantially more
than grass-roots monitoring, Muralidharan and Sundararaman (2011) where eﬀectively designed incentive schemes improved
student learning outcomes far more than unconditional grants, and Pomeranz (2015) in which pre-existing incentives for tax
compliance substitute for government audits.
    6
      This paper also contributes to the literature on public sector decentralization. We provide direct empirical evidence that
decentralized organizations have superior information, which is often assumed in the literature (Oates, 1999, 2005; Bardhan
and Mookherjee, 2006). Our evidence implies that within layers of an organization, information is relatively homogeneous,
while across organizational tiers it is relatively asymmetric. Our ﬁnding that the impact of decentralization on information
is mediated by management practices provides one potential explanation for the divergent ﬁndings of the decentralization
literature on service delivery (Bardhan, 2002; Crawford and Hartmann, 2008).


                                                               6
in hierarchy. Section 4 outlines our experimental design and presents results. Section 5 concludes.


2     Measuring Information

2.1   Institutional Background

Ethiopia is Africa’s second most populous country, home to 100 million people or 10% of the population of
Sub-Saharan Africa. Its ethnically diverse population is dispersed across over a million square kilometers,
such that population characteristics vary widely across space.
    Like other developing countries, government expenditures represent a signiﬁcant fraction of GDP (18%)
and the public sector is a large employer of formal workers. Corruption is less prevalent than in most of
Sub-Saharan Africa, but government eﬀectiveness is seen as relatively weak overall. Never fully colonized,
Ethiopia’s government is organized as a developmental state, with a relatively vertical and rigid hierarchy.
    The state is organized in a federal system, with three major tiers of government: federal, regional, and
district (woreda). Each tier has its own staﬀ of civil servants who must deﬁne and reﬁne government policy
within their tier, direct budgetary and other resources, and manage the implementation of public services.
Though the speciﬁc mix and focus of these tasks diﬀers across tiers, each tier undertakes these tasks to
a degree, and must make choices informed by the information they have on citizen needs. For example,
roughly half of all public resources are spent by the Federal Government, with the rest being directed by
the regional governments or spent directly by district administrations (World Bank, 2016b). None of the
oﬃcials we study implements public services themselves, at any of the tiers, but rather administers public
policy and resources.
    Across sectors, the structure of the vertical hierarchy is relatively constant. Ministry organizations
focus more on agenda setting for the sector, and the development of governing policies, but also direct
resources to service providers and thus focus on managing implementation. Regional organizations develop
policies for their region in line with federal stipulations, and guide resources across districts. District
governments focus on service delivery, while translating policies into guidelines for their district. Each tier
therefore makes decisions on how to distribute resources depending on where they believe there is most
signiﬁcant need. This is true in both the development of policy, which dictates the areas of the sector that
should receive resources, as well as in implementation, which assigns resources geographically. Federal and
regional organizations co-ordinate closely in their development of policy and operational activities, and
can be seen as a coherent ‘centralized’ body that provides joint guidance to district governments.
    The relationship between central government agencies and district governments is such that many
basic decision rights lie with the district government. Day-to-day decisions made by members of a district



                                                      7
organization cannot be overruled by the central government. The relationship between central and district
government agencies therefore accords closely to the theoretical literature’s notion of de facto ‘agent
authority’. In contrast, within an organization a manager has the ability to overrule the agents she
employs on business decisions. By contrasting the diﬀerences in information between principals (central
tiers of government or managers) and agents (district governments or employees) under these two regimes
allows us to assess the impact of the nature of authority on information acquisition.
      We can also use the fact that diﬀerent organizations in the Ethiopian context are managed diﬀerently
(as will be shown below) to investigate the impact of environments that diﬀer in terms of how they value
the acquisition of information. As the culture of an organization becomes more empirical, we can interpret
this as an increase in the individual rewards for information use. We can thus test the impact of mediating
incentives on the likelihood of information acquisition.


2.2      Survey of Public Oﬃcials

The implementation of any of the three core tasks of public administration outlined above requires infor-
mation on the characteristics and thus needs of the citizens they serve. To elicit an understanding of the
information public oﬃcials have about their environment and to understand the incentive environments in
which they operate, between June and September 2016 we held face-to-face interviews with 1,831 public
administrators in 382 organizations across the three tiers of Ethiopia’s government. Table A1 lists the
organizations included in our study, and Figure A1 provides a map of the district governments we sampled.
Our analysis focuses on the agriculture, education, health, revenue, and trade sectors only. For each of
these ﬁve sectors, we sample the main ministry, all the corresponding regional oﬃces, and a geographically
representative sample of corresponding district oﬃces from across the country.7 Within each organization,
we interviewed senior managers and a representative sample of their staﬀ. We limit our scope to the
professional grades of technical and administrative oﬃcers, excluding grades that cover cleaners, drivers,
secretaries, etc.
      Table 1 provides descriptives for the 382 organizations we study. All oﬃcials work within a relatively
standard structure, with a manager overseeing levels of hierarchy below him within a clearly deﬁned
organizational structure. As is common, the more centralized an authority is, the larger it is in terms
of both the number of managers and the number of employees. Federal organizations also have a higher
number of employees per director than either regional or district government oﬃces.
      In terms of bureaucrat characteristics, around 20% are women (at managerial grades, 8% are women,
at non-managerial grades 24% are women); 82% of Ethiopian bureaucrats have some form of university
  7
      Our survey, and therefore analysis, necessarily excluded the majority of the Somali region for security reasons.



                                                               8
education, with 11% having a postgraduate degree (at managerial (non-managerial) grades, 17% (10%)
have a post-graduate degree). As in other state organizations, bureaucrats enjoy stable employment
once in service: the average bureaucrat has 13 years in service, with their average tenure in the current
organization being 7 years.
   Across tiers, bureaucrats are similar in that they are in their mid-thirties, have been in their current
post for roughly similar amounts of time, and have all worked in roughly 3 service organizations. However,
oﬃcials at regional organizations are more senior and more likely to be a manager. Oﬃcials at district
organizations are less likely to have a university education, and only 1% of district oﬃcials have a master’s
degree. They are also more likely to be male, with 18% of district oﬃcials female compared with 23% of
regional oﬃcials and 27% of federal employees.
   To develop and enumerate a questionnaire that was relevant for the Ethiopian setting, we worked
closely with the Ministry of Public Service and Human Resource Development and employed ex-civil
servants within our enumeration teams to facilitate navigation of the public service. As will be described
in more detail below, we used the survey to ask a range of questions about citizen characteristics and the
public service rules to capture the knowledge of public oﬃcials with regards their working environment.
Beyond our focus on collecting bureaucrat knowledge, we surveyed oﬃcials on their basic characteristics
and the nature of management practices utilized at their organization. The implementation of the survey
was successful across the organizations we visited, with 99.5% of public oﬃcials sampled agreeing to be
interviewed and 98.2% of interviews being classiﬁed by the enumerator as having gone ‘somewhat well’
(26.4%) or ‘very well’ (71.7%).
   In conjunction with the survey, we also collected data on the performance of public oﬃcials as assessed
in their annual appraisal. Each year, public oﬃcials are evaluated by their direct manager on the tasks
that they were expected to contribute to. For example, tasks might include ‘Monitor and provide support
to the [work] team preparing the budget’ and ‘Support the team to prepare soft and hard copy documents
of the budget’. Managers evaluate the quality of contributions bureaucrats make to the tasks they were
involved in and produce an overall ‘performance’ score. In addition to this performance-related score,
public oﬃcials are evaluated on their ‘attitude’ to work, which intends to measure their oﬃce behavior
and alignment to the organization (Abagisa, 2014; Tereda, 2014). For the year 2016, we collected the
performance, attitude, and total scores (which are a weighted average of performance and attitude scores)
for each oﬃcial from a subset of the organizations we visited for which they were available.




                                                     9
2.3    Subjective Assessments of Public Oﬃcials

We used the public oﬃcial’s survey to ask each oﬃcial direct questions on the socio-economic conditions
of the citizens they serve and on civil service rules that govern their workplace. Respondents had to
answer based on their existing knowledge of their constituencies and workplace rather than by referring
to external reference documents. The questions we asked are listed in Table A2 and relate to both
district characteristics and civil service rules. Those relating to socio-economic characteristics of citizens
correspond to the core variables of interest in the national ‘Growth and Transformation Plan’, and sector-
speciﬁc plans, such as the ‘Education Sector Development Program IV: 2010 - 2015’. Those relating to the
public service correspond to the rules outlined in the ‘Ethiopian Federal Civil Servants Proclamation No.
515/2007’. Thus, all the variables we asked about are fundamental quantities relating to the core work of
the public oﬃcials we interviewed. For example, the core mission of education oﬃcers is to implement the
education sector development programs, focusing on the topics it prioritizes.
    There are three sets of questions. The ﬁrst set asks basic demographic questions about districts such
as population and the proportion of the population that lives in rural areas. These questions are asked to
respondents across all sectors. The second part asks sector-speciﬁc questions enumerated to sector-oﬃcials
only. For example, respondents working in the education sector were asked the following questions about
the citizens in their constituency: “What do you think the primary (1-8) enrolment numbers are for [district
name] according to oﬃcial administrative data?”; “What do you think the primary (1-8) pupil-teacher ratio
is for [district name] according to oﬃcial administrative data?”; “What do you think the primary (1-8)
pupil-class ratio is for [district name] according to oﬃcial administrative data?”; and, “What do you think
the primary (1-8) pupils-per-school ratio is for [district name] according to oﬃcial administrative data?” 8
The third set ask questions about the rules of employment of the civil service, such as the number of hours
a public oﬃcial is expected to work in a regular working week, the annual leave allowance in days in the
ﬁrst year of service, and so on. These are once again asked to all oﬃcials. All indicators asked about were
continuous in nature.
    Since the three tiers of government serve constituencies of diﬀering sizes, we asked federal oﬃcials which
region they knew best, and then asked them questions about the 25th, median and 75th percentile districts
within that region by socio-economic indicator. Regional oﬃcials were asked about the corresponding 25th,
median and 75th percentile districts within the region where they worked. District oﬃcials were asked
about the local government in which they worked. Thus, all oﬃcials made statements about speciﬁc
districts in their constituency, and in particular for the federal oﬃcials, those they stated they knew best.9
  8
    District-level sector-speciﬁc information was not available for the Addis Ababa City Administration, Dire Dawa City
Administration and Harar Regional Administration.
  9
    The exact phrasing of questions depended on the tier at which the respondent worked. Within each indicator, fed-



                                                          10
    For each question, public oﬃcials were permitted to state, “Don’t Know”, though the enumerator did
not suggest this as a response. In 23% of questions, the respondent stated that they simply did not know
and could not provide a valid guess. This is in contrast to the fact that the quantities assessed are the
core variables required to make policy, planning and distribution decisions in accordance with the national
sector plans.


2.4    Objective Assessments of Citizen Characteristics

For each indicator, we have administrative or survey data that can benchmark the responses by oﬃcials.
The speciﬁc sources for each variable are outlined in Table A2 and are the latest Census, the latest available
service delivery data from the education and health management information systems, and the Woreda
and City Benchmarking Survey (WCBS). Each data set is created by sourcing information directly from
service providers and aggregating it at the center of government. The sources we use are generally regarded
as the highest quality large-scale measurement undertaken in Ethiopia on the topics outlined above. For
example, in education, the Education Management Information System is comprised of school-level data
that are processed through each tier of government, oﬃcials of which undertake quality checks to ensure
that the data are as accurate as possible. The WCBS data was collected in collaboration with the World
Bank and the Ministry of Public Service and Human Resource Development as a monitoring tool over ﬁve
separate waves of data collection. All of this data is available to the oﬃcials we study through the relevant
directorates of their tier of government at the minimum, and in some cases online.10
    We take the administrative and survey data as a proxy for the underlying state of the world. We use
these variables to benchmark the claims of public oﬃcials. For each indicator, we subtract the correspond-
ing value in the benchmark data from the relevant claims in our data set, such that a positive error is an
overestimate.
    Our core analysis aims to bring together the full set of claims in a uniﬁed framework. However,
individual errors are not comparable across indicators. Making an error of 0.5 in a district’s pupil-class
ratio is large for that indicator, but tiny for the population indicator. To jointly assess the scale of errors
across multiple indicators, we undertook a number of transformations of the raw error data. First, we took
absolute values of the errors so that deviations of diﬀerent signs are comparable.11 Second, we trimmed
eral/regional level respondents are asked about the value of the 25th percentile district in their (chosen) region, the 50th
percentile district in their (chosen) region, and the 75th percentile district in their (chosen) region.
  10
     One proxy of the quality of the benchmarking data is their stability over time. By assessing the extent to which indicators
vary within districts over data collection rounds, we aim to understand whether the data reﬂect a noisy collection process.
Table A2 presents summary statistics and tests for the presence of a random walk for each of the indicators used in the
analysis where panel data are available, and the Appendix describes the results. We ﬁnd evidence that the benchmarking
data is consistent over time.
  11
     The probability of a negative (as opposed to a positive) error in the data we use for our core analysis is 0.404, such
that the likelihood of over- and under-estimates is relatively equal. We have also transformed the errors to be percentages



                                                              11
the data at the 95th percentile. In a very small number of cases, errors were so large that it seemed
infeasible for even the most uninformed bureaucrat to have made such an error, and it looked likely to be
a data entry problem. Either way, the interpretation of the remaining data is based on the errors of 95%
of the claims of public oﬃcials in our study.
   Our core transformation is to create z-scores of each of the variables by subtracting the mean and
dividing by the standard deviation of all the errors related to a particular indicator. This scales the claims
by an empirical measure of a ‘normalized’ measure of error for that indicator. The interpretation of the
z-score is a unitless measure of how many standard deviations of the full distributions of error away from
the benchmark data an oﬃcial’s assessment is. Thus, it is a measure of whether an individual oﬃcial made
a ‘large’ error relative to the empirical distribution of errors we observe for that indicator. Comparisons
across indicators therefore take into account the empirical distribution of errors for each indicator. They
compare the scale of an oﬃcial’s error relative to the errors made by all other oﬃcials in our data set,
indicator by indicator.
   Deﬁning z-scores in this way assumes that the tails of the raw errors are roughly symmetric for each
indicator, such that an error 10% below and above the truth are equally likely. However, the distribution
of errors of a number of the indicators we study are right-skewed, such that we may be underestimating the
severity of an error in the sense that few people would make it. To test the robustness of our core measure
to this criticism, we deﬁne a ‘skewness-adjusted’ z-score, where we test for skewness of the distribution of a
raw indicator and replace skewed indicators with their log. We then take the z-score of these transformed
variables in a similar way to the above measure.
   Finally, we move away from distributional concerns by taking the ordinal rank of absolute errors within
an indicator and divide this by the maximum rank, thereby creating a ‘rank-proportion’ score. This can
be interpreted as the percentile of error made by an oﬃcial relative to other oﬃcials in our data set. The
Data Appendix provides further details on how the error variables we use in the analysis were created.


2.5    Measuring the Incentive Environment

To understand the incentive environments in which public oﬃcials in the organizations we are studying
make decisions on information absorption, we also collected data on management practices. Following
recent eﬀorts to collect data on the management practices of public administrations (Rasul and Rogger,
2018; Rasul et al., 2018) we used a public sector version of the World Management Survey (WMS; Bloom
and Van Reenen, 2007; Bloom et al., 2012) to elicit measures of the management practices under which
public oﬃcials operate.
of the underlying benchmark data, and then created the various measures outlined below. Taking this approach does not
qualitatively change our core results.


                                                         12
       The WMS evaluation tool elicits management practices through a semi-structured interview covering
7 topics: ﬂexibility, incentives, monitoring, roles, staﬀ involvement, staﬃng, and targeting. Table A3
details each of the 19 management-related questions, by topic, as well as the 1-5 scoring grid used by our
enumerators for each question. These questions provide a holistic assessment of practices related to the
topic.12 We take the ‘Management Practice: Monitoring’ index as a proxy for the organization’s value
of the information acquired by oﬃcials and the ‘Management Practice: Other’ index, which aggregates
the non-monitoring sub-topics of the WMS, as a measure of alignment.13 The Appendix provides further
details on how we constructed our management measures.
       Figure A2 plots the distributions of the ‘Management Practice: Monitoring’ and ‘Management Practice:
Other’ indices across the organizations we use in our core regressions.14 Both indices exhibit substantial
variation, indicating that de facto management practices vary substantially across the Ethiopian Public
Service. For practices related to the organization’s value of the information acquired by oﬃcials, the
75th percentile organization has a ‘Management Practice: Monitoring’ score that is 141% higher than
the 25th percentile organization. For practices related to the alignment of preferences of agents within
an organization, the 75th percentile organization has a ‘Management Practices: Other’ score that is
135% higher than the 25th percentile organization. This variation occurs despite organizations operating
under the same broad political environment, the same federal administration, sharing the same national
history, and working under the same governing rules over budget allocations, wages, non-wage beneﬁts,
promotion schedules and other contracting arrangements. This variation across organizations is consistent
with evidence from other management surveys of the public sector (Bloom et al., 2015; Rasul and Rogger,
2018; Rasul et al., 2018).
  12
      To provide a sense of the holistic (rather than speciﬁc) nature of these questions, we go through one example: a question
relating to management practices relating to monitoring was, “In what kind of ways does your Directorate track how well
it is delivering services? Can you give me an example?” Enumerators could then score responses on a continuous 1-5 scale,
where for indication the scoring grid described a score of one as corresponding to circumstances where the “Directorate does
not track performance”; a score of three corresponded to, “Directorate tracks a number of performance indicators. These
are seen and reviewed by senior management only”; and a score of ﬁve corresponded to, “Full set of indicators are tracked
formally and continuously. Reviews are conducted regularly and involve representatives of all directorate staﬀ groups. The
results of the review are formally communicated to all staﬀ.”
   13
      These topics measure multiple aspects of management including staﬀ involvement and incentives, all of which could be
said to be means of aligning the preferences of agents to those of principals. For example, what are performance incentives
but a mechanism for aligning agent action to the desires of principals?
   14
      The ‘Monitoring’, and ‘Other’ indices are positively correlated with each other, with a coeﬃcient of 0.62. Importantly,
the partial correlations of these two dimensions of management can still be separately estimated from each other.




                                                              13
3         Descriptive Evidence on Information in Hierarchy

3.1         The Errors of Oﬃcialdom

The left-hand panel of Figure 1A provides histograms of both the benchmark data and public oﬃcials’
claims regarding the population of the districts that they serve. The histograms are divergent at many
points in the distribution, indicating that in general public oﬃcials are making errors about the size of
the underlying populations they are serving. The right-hand panel of Figure 1A provides a histogram and
density plot of the scale of errors made by public oﬃcials in their estimates of the populations they serve.
         The distribution of errors in Figure 1A is perhaps as would be expected. On this speciﬁc indicator, the
average error is close to 0, and the distribution is close to being normally distributed. However, our data
allow us to provide estimates of the scale of the standard deviation in the errors of public oﬃcials. The
dashed vertical line represents the median population of a district in Ethiopia (120,398) to give a sense of
the magnitude of the errors for this indicator. Many oﬃcials’ claims are relatively accurate; 21.5 percent
of public oﬃcials guess within 20% of the benchmark data. On the other hand, the scale of error of a
substantial fraction of oﬃcials’ claims is large; 16% of claims by oﬃcials over or underestimate the size
of the claimed district by more than the population of the median district. 37% of oﬃcials overestimate
or underestimate the population of the district by more than half the population of the median district.
And 47% of oﬃcials claim that their district’s population is 50% bigger or smaller than it is.15
         The scale of errors is similar across the distribution of district populations, implying that public oﬃcials
are indeed adjusting their estimates towards the underlying state of nature. The left-hand panel of Figure
1B exhibits the scale of errors made by public oﬃcials in their claims about districts that have the 25th
percentile smallest population in their region, the median, and the 75th percentile. The right-hand panel
of Figure 1B exhibits the corresponding proportional errors. That is, the proportion of the benchmark
population made up by the claim. In each case, over 50% of oﬃcials claim that they are serving a
population that is at least 50% smaller or 50% larger than it is. Thus, across the full range of district sizes
in Ethiopia, a sizeable fraction of public oﬃcials is making substantial errors as to the number of citizens
they are serving.
         In Figure 1C, we present a series of thought experiments that allow us to assess the economic implica-
tions of the scale of the errors we observe. Suppose that federal and regional public oﬃcials were tasked
to distribute public resources evenly across the district populations in their region of interest. The panels
of Figure 1C present the de facto weights the oﬃcials we study unwittingly embed in their distribution of
    15
    A simple test of distributional knowledge is to ask federal and regional oﬃcials whether they can correctly rank the
25th, 50th and 75th percentile districts in their region of expertise in terms of population. In our data, only 18% of oﬃcials
state the correct ordering for their region (this is true for both federal and regional oﬃcials separately). 47% of these central
oﬃcials cannot accurately state which of the 25th and 75th percentile districts in their region has the larger population.



                                                               14
resources across individual citizens by dividing the perceived population by the actual population. A value
of 1, therefore, signiﬁes that the civil servants assigned a weight that accurately reﬂects the population
data. To the far right of the top panel we see that one oﬃcial in our data would mistakenly weight the
population of one district 3.5 times as high as the actual population would justify. And 14% of oﬃcials
would weight the median district in their region at least 1.5 times more than its population justiﬁed.
Relative to an equitable distribution of government resources across constituents, central oﬃcials would
over disburse by 30% to districts at the 75th percentile of the distribution of errors, and under disburse
by 12% to those at the 25th percentile of errors. The middle panel restricts this analysis to managers only
and indicates that individual civil servants and their managers both make economically meaningful errors
on this simplest of public resource distribution tasks. Compounding this with other errors, such as the
mis-measurement of primary pupil-teacher ratios, will distort the distribution of resources even further.
Even in the best case scenario that capitalizes on all the information in an organization shown in the
bottom panel of Figure 1C, such compounded errors would create substantial distortions in public sector
choices.
   The distribution of errors among other variables is similarly large. The estimates of 16% (23%) of
public oﬃcials regarding the characteristics of their constituents are at least a standard deviation (half
a standard deviation) away from the objective benchmark data we collect. The mean absolute error in
education bureaucrats’ estimates of primary enrollment numbers is 76% of the true enrollment ﬁgures.
Across all of the education variables, the mean absolute error is 51% of the benchmark data. The mean
error in estimates of the proportion of pregnant women who attended ANC4+ during the current pregnancy
(the ‘antenatal care rate’) was 38% of the benchmark data. Agriculture oﬃcials overestimate the number
of hectares in their district that are recorded as used for agricultural purposes by almost a factor of 2.
As noted above, these magnitudes imply substantial misallocation of resources across space and subject
matter. If bureaucrats think that antenatal care is a larger issue than it is, they will devise policies
that prioritize corresponding investments beyond those that are justiﬁed by the underlying state of the
citizenry.
   Figure 2 provides cumulative distribution functions of the three aggregated measures of error that we
use in this study. Figure 2A presents the distribution of the unadjusted z-scores, with separate plots for
Ethiopia’s three tiers of government. Since these represent z-scores of absolute errors, a lower z-score
implies a more accurate claim. The ﬁgure highlights substantial variation in errors across government
that echoes the variation in the raw indicators described above and we observe oﬃcials making the full
distribution of claims. Previewing later results, we see that oﬃcials at the district tier of government make
lower errors at almost every point in the distribution. A Kolmogorov-Smirnov test of distributional equality



                                                     15
is rejected with a p-value of 0.00. Regional and federal organizations have almost identical distributions
of errors, though regional organizations are slightly more prone to outliers.
      Figure 2B plots the distribution of z-scores where adjustments are made to skewed underlying distri-
butions when necessary. Conditional on diﬀerences in the ranges of the distribution produced by the two
diﬀerent methods, the graph is qualitatively similar to that of the unadjusted z-scores. The adjusted scores
have a more even distribution of extreme values, as would be expected. Figure 2C plots the rank-proportion
transformation, and while district errors continue to be substantially smaller than those at higher tiers,
some distinctions arise between the regional and federal government levels. Overall, along the margin of
decentralization across tiers, the three measurement methods provide highly consistent messages.


3.2      What is the Appropriate Conception of Information in the Public Sector?

These errors in bureaucratic beliefs matter because they are used as the basis of public policy decisions.
In our representative survey of public oﬃcials, we asked public servants, “What are the three most impor-
tant sources of information you use to ﬁnd out about the state of service delivery in your jurisdiction?”
along the following margins: ‘Informal interactions e.g. through co-workers’, ‘Management Information
Systems’, ‘Formal ﬁeld visits’, ‘Reports from the frontline’, ‘Information from higher-tier institutions’,
‘Formal citizen/user interactions’, or ‘Media sources’. Oﬃcials were asked to rank their top three sources
or specify that others are more important.16
      In our setting, information is dominantly sourced from inside the public service rather than externally
from the media. Only 8.93% of public oﬃcials in our survey responded that media sources were one of
their top three sources of information, and 2.12% of oﬃcials selected it as their top source of information.
This ﬁnding was consistent across tiers, with 8.54% of federal, 8.56% of regional and 9.18% of district
oﬃcials stating media was one of their top-3 information sources, across sectors and across grades. That
public oﬃcials are heavily reliant on information produced by the public service is consistent with the
ﬁndings of AidData’s ‘Decoding Data Use’ project (Masaki et al., 2017). Similarly, a review of Ethiopia’s
‘Sector Development Program IV: 2010-2015’ documents for the ﬁve sectors studied in this paper ﬁnds
almost no references to external sources.
      The most frequently cited source of information was ‘Formal ﬁeld visits’, with 63% of oﬃcials stating
that this was a key source of information. Discussions with frontline colleagues, and informal interactions
with colleagues were the second and third most cited source of information, with 51.9% and 45.9% of
oﬃcials stating their signiﬁcance respectively. A similar pattern is observed restricting the selections to
the top cited source of information for each individual: formal ﬁeld visits are most frequently cited (22.3%),
 16
      Only 4% of oﬃcials stated that one of their top three sources of information was not included in our survey options.



                                                              16
followed by informal interactions with colleagues (19.6%) and then reports from the frontline (17.2%).
       These responses indicate that the dominant form of information used for decision making in the
Ethiopian public sector is in terms of knowledge transmitted or gained through personal experience,
widely known as the ‘tacit’ knowledge of individuals. Together, information gained through the direct
experience of the individual - from ﬁeld visits or interacting with colleagues or citizens - is one of the top
three sources of information collection for 90% of the oﬃcials in our data. Only 12.8% of oﬃcials state
that they use management information systems (MIS) as their primary source of information. Field visits
and informal interactions are 3 times more likely to be in the top 3 most important sources of information
than MIS, and 10 times more likely than external media sources.17 These ﬁndings for Ethiopia’s public
service are consistent with wider ﬁndings from the World Development Report 2016 that highlighted how
public oﬃcials across the developing world fail to capitalize on digital databases available to them.18
       This lack of use of MIS does not seem to be due to a lack of skills. Our survey asked the extent to
which oﬃcials agreed with the statement, “I possess the necessary skills to make best use of the technology
that is available to me.” Two-thirds of respondents stated that they ‘slightly’ or ‘strongly’ agreed with
the statement. Similarly, education indicators explain less than 1% of the variation in the probability that
oﬃcials state that they use informal interactions in their top 3 most important sources of information.
The lack of use of MIS also does not seem to be due to variation in facilities such as hours of electricity an
organization typically receives, the quality of the phone network or internet access. An aggregate index
of the quality of facilities at the organizations we study also explains less than 1% of the variation in the
likelihood of relying on tacit knowledge.19 Rather, the culture of information use leans strongly towards
a reliance on tacit knowledge.


3.3      Basic Descriptives on the Determinants of Information

Given the importance of the tacit beliefs of bureaucrats measured by our survey, we now turn to an
investigation of their determinants. Our data allows us to provide broad descriptive evidence on the
  17
      The reliance on tacit knowledge is in contrast to the fact that many public oﬃcials believe relevant information is
available from other sources. We asked oﬃcials how long it would take to receive statistical data as an input to their work
from the relevant directorate, and the median response was 1 day (the mean was 1.8 days and the 90th percentile was 2
days). Averaging bureaucrat responses to the organization level, the 90th percentile slowest organization delivered the data
within 3 days.
   18
      Many public oﬃcials believe that the tacit information they have is suﬃcient for the successful implementation of their
daily work. Our respondents stated that in 70.9% of the projects they worked on, “I have necessary information regarding
the current state of service delivery in my jurisdiction/my area of work”. This was consistent across tiers (federal, 73.6%;
regional, 72.9%; district, 69.4%), sectors (Agriculture, 71.1%; Education, 75.3%; Health, 74.3%; Revenue, 66.1%; Trade,
67.6%) and grades (Grades 1-5, 71.1%; Grades 6+, 70.7%).
   19
      The index of facilities is an organization-level average of z-scores of the facilities questions asked to managers only:
how many hours of electricity an organization receives per typical working day, the state of the phone network, the extent
of internet access, the access staﬀ have to computers, the access staﬀ have to vehicles, and the staﬀ’s ability to write
memos/spreadsheets/powerpoint.




                                                             17
determinants of superior information in public hierarchies. Such descriptive evidence provides a broad
framework in which to understand the experimental analysis.
   We undertake an ANOVA analysis of our error data using variables related to the hierarchical struc-
ture of the bureaucracy and the marginal costs and beneﬁts of information acquisition. The results are
presented in Table A4 and described in the Appendix. We ﬁnd that structural variables, such as the tier
of government, the speciﬁc indicator we test on, and organizational incentives explain the dominant pro-
portion of variation in the errors. Individual characteristics of public oﬃcials and environmental features,
such as speciﬁc features of the population and the proximity to local constituents, have little explanatory
power. We therefore focus on the importance of hierarchical structure and organizational incentives in
what follows.


The Impact of Formal Authority

If a manager has de facto authority over an employee, this limits the agent’s incentive to invest in in-
formation acquisition due to concerns that she will be overruled and the temptation to free-ride on the
principal’s eﬀorts. Relationships in which the agent has formal authority should be characterized by their
having superior information on local conditions. In the Ethiopian government, this would imply claims
by district oﬃcials will be more accurate than those by central government oﬃcials, whilst we should ﬁnd
limited diﬀerences across managers and non-managers within individual organizations.
   Our main estimates of the impact of hierarchy on information therefore uses an OLS regression of the
following form,


                        yc = γ1 1 {District = 1} + γ2 1 {Manager = 1} + Xc +         c                     (1)

   where yc is a dependent variable relating to the scale of the error of the claim, such as the z-score of the
absolute error, 1 {District = 1} is a dummy that takes the value 1 if the claim is made by a district oﬃcial,
1 {Manager = 1} is a dummy that takes the value 1 if the oﬃcer making the claim is a manager, and
Xc is a set of sector, district, indicator and percentile ﬁxed eﬀects, individual control variables, marginal
cost and noise controls. Individual-level controls are tenure in the civil service of the respondent making
the claim; tenure in position; tenure in organization; an indicator for whether the respondent is female;
an indicator for whether the respondent holds an undergraduate degree; and, an indicator for whether the
respondent holds a master’s degree. Given the proximity of the federal and regional errors, as described
above, we aggregate these into a single ‘centralized’ tier. When we run the regressions with all three tiers
of government, the federal coeﬃcient is not signiﬁcantly diﬀerent to the regional at the usual levels, as




                                                     18
will be documented below.20
       The regressions are at the claim level, so comparisons can be made across individual percentile-indicator
pairings, providing the most precise comparison we can oﬀer. In our preferred speciﬁcation, the variation
that identiﬁes the core coeﬃcients is that between two sets of claims, made about the same indicator,
at the same percentile of the corresponding distribution, in the same sector, about the same district
by individuals at the two levels of hierarchy, conditional on their individual characteristics. Thus, our
speciﬁcation ensures a relatively tight comparison between levels of the hierarchy. We restrict our sample
to those claims that are made about the same districts to ensure the variation in our regressions arises
only from coherent comparisons. That means the claims used for comparison are made by 594 federal
and regional government oﬃcials, and 297 district oﬃcials. Since there may be correlations in the sectoral
information about a particular district, we cluster at this level throughout our analysis.
       Table 2 presents the results. In columns 1 and 2, the dependent variable is a binary that takes the
value 1 if the individual oﬃcial stated that they did not know the answer to a question, and thus could not
make a claim. We see in column 1 that district oﬃcials are 17 percentage points less likely to answer ‘Don’t
Know’, while there is almost no diﬀerence between managers and non-managers in readiness to provide
an estimate. Controlling for sector, district, indicator, percentile and individual controls in column 2 of
Table 2 does not change these results. This implies that if you were to enter oﬃces at diﬀerent tiers of
government, you would face a richer information environment at the district level even conditional on the
level of errors. The error results should therefore be interpreted as ‘the perceived state of knowledge within
an oﬃce’, and imply they represent a lower bound on the diﬀerence in knowledge between the two tiers.
       Columns 3 to 5 of Table 2 present unconditional versions of (1) where the dependent variable is the
z-score of errors across socio-economic variables. In column 3 we include a district government dummy
only. We see that across all claims those made by district oﬃcials have signiﬁcantly lower errors than
those of central oﬃcials. In contrast, column 4 includes only a manager dummy and indicates that there
is no diﬀerence in the magnitude of errors made by managers and their employees. In column 5 we include
both dummies, and ﬁnd that the results do not change.
       This pattern of results is reﬂected in regressions using variation from within the sector (column 6),
constraining comparisons to be based on claims about a common district (column 7), focused on common
indicators (column 8) and on common percentiles of those indicators (column 9). Column 10 presents
our preferred speciﬁcation, where we also condition on characteristics of the individual oﬃcial making the
claim. It presents some of the most precise empirical evidence we know conﬁrming a classic assumption
  20
     The fact that there is little diﬀerence between the federal and regional tiers is also evidence against physical distance or
remoteness playing a critical role in the acquisition of information in this setting (Agarwal and Hauswald, 2010; Landier et
al., 2007).




                                                               19
in organizational theory; that agents have superior information about local conditions. However, in these
data we ﬁnd that agents with formal authority (those acting in district organizations) have superior
information, while those who do not (those acting under a principal within the same oﬃce) have no better
information than their seniors.21 The superior information of agents is a product of a particular form
of decentralization rather than a rationale for it. In terms of the geography of information in the public
sector, these results imply that information is spread relatively evenly across layers of the hierarchy within
an oﬃce, but that signiﬁcant asymmetries occur across oﬃces.22
       The coeﬃcient on the district government binary is -0.33, implying that claims made by district oﬃcials
are on average 0.33 standard deviations closer to the benchmark data than central governments across the
z-scores of variables related to citizen characteristics. As can be seen by comparison to the coeﬃcients
in Table A4, such institutional features dwarf the impact of greater tenure in the service or increased
formal education. The coeﬃcient on district government is also an order of magnitude more important
than if an MIS system exists within the corresponding organization. If we remove district ﬁxed eﬀects to
allow for estimation of the impact of district poverty rates on errors, we see that the coeﬃcient on the
poverty rate (measured from 0 to 100) is 0.008 with a p-value of 0.05. Thus, oﬃcials make greater errors in
poorer districts, as we might expect. A one percentage-point increase in poverty is associated with a 0.008
standard deviation increase in the absolute error. The size of the coeﬃcient on district government in
Table 2 is 41 times larger in absolute terms, such that the quality of information in district organizations
is akin to estimating over the 25th percentile of poor districts rather than the 66th.23
       Our ﬁndings provide little support to theories in which decentralized agents (such as local governments
in a multi-tier system) are assumed to have superior information due to their proximity to local conditions.
As Oates (1999) states, “individual local governments are presumably much closer to the people and geog-
raphy of their respective jurisdictions; they possess knowledge of both local preferences and cost conditions
that a central agency is unlikely to have”. Bardhan (2002) argues that the contracting environment of local
governments incentivizes information acquisition through political accountability or yardstick competition,
and that the cost of access to information is lower relative to central level governments.24 Our evidence is
  21
      In our preferred speciﬁcation, the interaction between district government and manager is not signiﬁcant at the usual
levels, with a coeﬃcient of -0.07 and a p-value of 0.58 (the coeﬃcient on district government is -0.307 with a p-value of 0.004
and the coeﬃcient on manager is 0.053 with a p-value of 0.52). Thus, at both central and district tiers of government, there
is a common information environment within the organization.
   22
      This assertion is supported from statements by oﬃcials in our survey. Only 4% of central level oﬃcials (1% of federal
and 6.5% of regional) select ‘formal interactions with lower-tier institutions’ as their top source of information.
   23
      To assess the robustness of these core results to our methods of measurement, we present a series of robustness checks
in Table A5 and discuss them in the Appendix.
   24
      It is not necessarily the case that the local government should have superior information. Under the revelation principle,
it is feasible that the central government can perfectly extract the information of the local government (Myerson, 1981;
Mookherjee, 2006). As Cremer et al. (1994) argues, in acquiring information “there is no reason why central government
cannot use any of the techniques available to local government”, such that it must be an incentive issue.




                                                              20
much more strongly in favor of theories in which superior information at decentralized levels is a result of
greater control rights.25


Management Matters

The impact of those control rights can be mitigated by organizational incentives (Aghion and Tirole, 1997).
Recent empirical evidence has shown that management aﬀects the distribution of assets across ﬁrms such
as human capital (Bloom and Van Reenen, 2007; Qian et al., 2015; Bender et al., 2018) and physical capital
(Bloom and Van Reenen, 2010). It therefore seems plausible that it would have substantive impacts on
the distribution of information.
       The diversity in management practices across Ethiopia’s public organizations, and speciﬁcally in terms
of how highly information is valued as an organizational asset, can be interpreted as variation in the
rewards to oﬃcials of acquiring information. Organizations that value the use of information, captured by
the ‘Management Practices: Monitoring’ index, create a high relative reward for obtaining information.
In environments in which organizations value information, captured in the ‘Monitoring’ index, we should
expect to see lower errors. Conversely, management practices that raise the alignment of the principal and
agents’ interests may lead to a strategic substitution of eﬀort in acquiring information.
       We can extend our analysis in (1) to include measures of management practices and assess the consis-
tency of these ideas with our data. To do so, we broaden the variation across which we make comparisons
by switching from district ﬁxed eﬀects to regional ﬁxed eﬀects. Our reasoning is that since management
practices vary at the organization-level, within a district we constrain the variation to be across 5 or-
ganizations. By extending our estimating variation to the region, we are able to capitalize on variation
across an average of 122 organizations while still constraining the comparisons of errors within a restricted
geographic and institutional environment. We also follow standard practice in studies of the World Man-
agement Survey indices, and include ‘noise controls’, which aim to soak up aspects of the enumeration
of the management survey.26 Finally, to focus the analysis on the beneﬁts of acquiring information, we
condition on proxies of the marginal cost of that acquisition.27 To ensure the results of Table 2 tolerate
  25
     Whilst we have focussed on oﬃcial’s beliefs about the populations they serve, we also collected data on their beliefs over
the rules of the public service. Constituency features could be regarded as ‘soft’ information, whilst public service rules,
written in the governing documents of all organizations, could be regarded as ‘hard’ information. Indeed, as laid out in
the Appendix, we ﬁnd that centralized oﬃcials have superior information on this hard information in line with theoretical
predictions (Myerson, 1981; Mookherjee, 2006; Dessein, 2002).
  26
     Noise controls are composed of the time of day of the survey, day-of-survey ﬁxed eﬀects, enumerator ﬁxed eﬀects, an
indicator for the enumerator’s subjective assessment of the quality of the interview, an indicator for the duration of the
interview (decile ﬁxed eﬀects), controls for the average tenure and gender of the managers interviewed.
  27
     Marginal cost variables are the proportion of managers in the organization that state that a management information
system is in place; the organization average response to “In what proportion of projects would you say information ﬂows
eﬀectively between individuals in the same organization?”; the organization average response to “In what proportion of
projects would you say information ﬂows eﬀectively between organizations?”; the organization average response of managers
to “How many personnel do you manage?”; the organization average response of employees to “How many people would you


                                                              21
these changes, we present our preferred speciﬁcation with these modiﬁcations in column 8 of Table A5.
We see that the coeﬃcient is -0.28, still signiﬁcant at the usual levels though now at the 10% level.
    Table 3 presents the results. In Column 1, we include the management indices, the manager dummy,
and only those controls equivalent to our preferred speciﬁcation in Table 2. We see that the coeﬃcients on
the management indices accord with theory exactly. Increasing the value an organization places on the use
of information, proxied by the ‘Management Practices: Monitoring’ index, decreases errors substantially.
The corresponding coeﬃcient is -0.17, and the p-value 0.014. Similarly, the coeﬃcient on the ‘Management
Practices: Other’ index is 0.28, and has a p-value of 0.005. Managers continue to have information
equivalent to their employees. These results are slightly strengthened by the inclusion of noise controls
and proxies of the marginal cost of information acquisition, as described in column 2. Combining the
indicator of de facto authority, the dummy for district government, and the management practices that
proxy incentives leaves the qualitative results consistent with the theoretical literature. Once again, these
results indicate the primacy of incentives for determining the distribution of information in the public
sector.
    In column 4 of Table 3, we undertake a ﬁrst assessment of whether the impact of de facto author-
ity varies by the nature of management practices. We include an interaction between the monitoring
management practice index and the district government binary. It is negative, and large in magnitude
relative to the baseline impact of de facto authority, but not signiﬁcant at the usual levels, with a p-value
of 0.17. However, we take this heterogeneity analysis further in Figure 4 by plotting the distribution of
coeﬃcients of the district government binary as we vary the decile of management practice. The ﬁgure
makes clearer that the impact of de facto authority is clearly mediated by incentives. At the lowest levels
of management practices observed in our setting, where incentives to acquire information are weakest,
providing agents with de facto authority has zero impact on their relative magnitude of errors. At the
highest level of management practices, the corresponding coeﬃcient is -0.66 with a p-value of 0.04. These
results imply that unconditional decentralization, without eﬀective corresponding management practices
or more broadly incentives for information acquisition, does not improve the information environment.
This ﬁnding could help reconcile the divergent ﬁndings on the impact of decentralization on service deliv-
ery (documented in surveys such as Bardhan (2002) or Crawford and Hartmann (2008)), as management
practices are typically not measured in such studies.28
say regularly give you tasks as part of your formal work duties?”; and, the additional alignment index that complements our
main measure.
   28
      In the setting of our study, we do ﬁnd evidence that management practices vary across tiers. Table A8 reports on
organization-level regressions where our management indices are the dependent variables, the district government binary
is the explanatory variable of interest, and a range of controls are presented. We ﬁnd that once diﬀerences in individual
characteristics and sector ﬁxed eﬀects are conditioned on, the coeﬃcient on district government is large (0.41) and signiﬁcant
at the 5% level, indicating that on average, district governments have more of a culture of monitoring. This ﬁnding is
strengthened once we condition on diﬀerences in the marginal cost of information acquisition. We also ﬁnd weaker evidence


                                                             22
    Returning to Table 3, we assess whether management practices have a diﬀerential eﬀect by whether
an oﬃcial is a manager. Organizational monitoring practices have slightly stronger eﬀects for managers,
but the coeﬃcient for non-managers is still negative and signiﬁcant at the 5% level, while the coeﬃcient
relating to the interaction is insigniﬁcant at the usual levels.
    We can also assess the baseline impacts of management practices that align principals and agents.
When the objectives of the principal and agent are closely aligned, this increases the incentives to ‘free-ride’,
as each trusts the recommendations of the other, reducing overall investments in information acquisition
(Aghion and Tirole, 1997; Myatt and Wallace, 2012). In line with this prediction, the dominant eﬀect of
alignment in our setting, as measured by the ‘Management Practices: Other’ variable, is to increase errors
across the columns in Table 3. The coeﬃcient in column 3 is 0.32 and signiﬁcant at the 1% level, thus large
compared with our other coeﬃcients of interest.29 We do not ﬁnd any evidence of these results varying
across the hierarchy. The coeﬃcient on an interaction of the manager binary and the ‘Other’ index is 0.00,
with a p-value of 0.97. Our results therefore imply that increases in alignment have a broadly negative
impact on the incentive to acquire information as oﬃcials aim to free-ride on their colleagues eﬀorts.


3.4    Information, Promotion, and Service Delivery

Before turning to our experiment, we are able to provide descriptive evidence on the impact of information
acquisition on an individual oﬃcial’s career trajectory. By regressing the data we have collected on
individual appraisal scores, we can assess how the magnitude of errors an individual is likely to make in their
claims about citizen characteristics correlate with their career progress. To ensure that we take account of
diﬀerences in scoring systems across organizations, we convert appraisal scores into organization-speciﬁc
z-scores, and cluster at the district government-sector level. Since our interest is beyond comparisons of
common claims across the hierarchy, we expand our sample in this regression to all claims at organizations
at which we have appraisal data, leading to an expansion in the number of claims we estimate over to
3,963 claims at 135 organizations.
    Table A9 presents the results. Column 1 indicates that unconditionally, a standard deviation increase in
absolute error is associated with a reduction in appraisal scores of 0.02 of a standard deviation. Controlling
for tier (column 2), sector (column 3), indicator (column 4), percentile (column 5), and region (column
6) leads to a coeﬃcient of -0.028, which is signiﬁcant at the 5% level. This is true if one controls for
individual controls (column 7), though the characteristics of oﬃcials account for some of the variation,
and for diﬀerences in organizational characteristics (columns 8 and 9).30 Thus, information acquisition is
for other management practices.
   29
      We do not ﬁnd equivalent baseline eﬀects using our secondary measure of alignment, which directly asks oﬃcials how
aligned they feel. The corresponding coeﬃcient is 0 and insigniﬁcant at the usual levels.
   30
      In terms of the impact of basic characteristics on appraisal scores, managers perform signiﬁcantly better on the evalua-


                                                             23
positively correlated with incentives in the public sector, consistent with the results in the previous section.
Interestingly, admitting to not knowing the characteristics of the citizenry does not have signiﬁcant impacts
on an oﬃcial’s appraisal scores in any of the speciﬁcations presented.
    Beyond career concerns, information may have impacts on the quality of public services delivered by
the bureaucrats we study. We therefore use z-scores of the underlying service delivery indicators that we
use to build the error indicators as a dependent variable. Table A10 presents the results of a regression of
absolute error (variously measured) on these outcomes. Again, since our interest is beyond comparisons
of common claims to a focus on service delivery, and to ensure some degree of homogeneity in incentive
environment, we expand our sample in this regression to all claims at district organizations for which we
have corresponding service delivery data. This leads to expansion in the number of claims we estimate
over to 2,853 claims at 276 organizations.
    Throughout the speciﬁcations in Table A10, the absolute error does not have clear impacts on service
delivery outcomes. Once suﬃcient controls have been conditioned on, oﬃcials stating that they ‘Don’t
Know’ has a negative impact on service delivery outcomes, of the order of 0.2 standard deviations of
the average service delivery indicator z-score. This is akin to an education oﬃcial stating ‘Don’t Know’
leading to a reduction in the enrollment rate of 9.2 percentage points, or in health of a reduction in the
antenatal care rate of 6 percentage points. More research is needed on the precise impact of bureaucratic
information on service delivery outcomes. In our setting, Somani (2018) ﬁnds evidence that the histor-
ical implementation of the education management information system increased enrollment rates by 2.7
percentage points. Similarly, other studies of improvements in bureaucratic information have tended to
ﬁnd positive eﬀects on service delivery (for example, on public health worker attendance in Callen et al.
(2018) and Dhaliwal and Hanna (2017), on the implementation of India’s rural employment guarantee in
Muralidharan et al. (2016), and on targeting interventions in Dal Bó et al. (2018)).


4     Experimental Evidence on Information in Hierarchy

4.1    Experimental Design

Our descriptive analysis has implied the importance of organizational incentives in determining the ac-
curacy of bureaucratic beliefs, and in particular those that increase the marginal beneﬁts of information
acquisition. We capitalized on the relative feasibility of an experiment in this setting to test the impacts of
a reduction in marginal costs on the accuracy of bureaucratic beliefs. This allows us to separately identify
the eﬀect of reducing access costs from other changes that come with delegating authority or decentralising
tions, while females achieve signﬁcantly lower scores at the usual levels. Education and tenure are not signiﬁcantly associated
with the evaluations at the usual levels.



                                                              24
activities. We aimed to design an intervention that would symmetrically reduce principal and agent costs
of information acquisition dramatically. Even in the presence of free-riding, a large symmetric reduction in
costs of information acquisition can increase the general accuracy of agent beliefs in a hierarchy (Aghion
and Tirole, 1997; Garicano, 2000; Gabaix et al., 2006).31
       Our intervention is based on a common approach to improving the information of public oﬃcials - the
evidence brieﬁng. Such brieﬁngs are now commonplace as a tool of government or research organizations
whose aim is to inform public oﬃcials. By condensing research ﬁndings, evaluations or survey data
into focused communications, evidence brieﬁngs aim to inform the decision-making of public oﬃcials
by reducing the cost of interacting with the raw analysis.32 Government circulars have been widely used
throughout history to provide information to a body of public oﬃcials. Recently, forms of evidence brieﬁng
have begun to be routinely produced by economics research institutions as diverse as the Bipartisan Policy
Center, the Centre for Economic Policy Research, the Institute for Fiscal Studies, JPAL/IPA, and the
World Bank Group. The use of evidence brieﬁngs is part of a broader eﬀort to make public policy
‘evidence-based’ by informing the beliefs of public oﬃcials (Smith, 1996; Pawson, 2006, Parkhurst, 2017).
       Prior to the enumeration of the survey, a random sub-sample of the organizations we study were sent
the oﬃcial administrative data corresponding to our policy variables of interest as a government circular.
A set of federal organizations received data for all regions for all indicators, highlighting the median district
for each indicator for each region. Similarly, a random sub-sample of district jurisdictions were sent the
oﬃcial administrative data for their whole region for all indicators, also highlighting the median district
for each indicator. Since the information package to the district-level jurisdictions had to pass through
the regional oﬃces, this meant that all regional oﬃcers were treated. Table A11 presents a comparison
of treatment and control groups across key characteristics of individual oﬃcials and the organizations
in which they work. Treatment and control groups look balanced across individual and organizational
characteristics, diﬀering on a single variable at the usual levels of signiﬁcance (years in position).
       For the treated district jurisdictions, the information package was sent to the Regional Bureau of Public
Service with instructions to pass on the data to the randomly selected districts. This primarily occurred
through the Woreda Oﬃce of Public Service, which then disseminated the information to the oﬃces within
the district jurisdiction. Since all regions were asked to communicate with the district administrations
of the arrival of the survey team in advance (to prepare the logistics, in terms of oﬃce space, in terms
  31
     Reducing the marginal cost for the principal (agent) only will increase the investment of the principal (agent) but decrease
the investment of the agent (principal) due to the strategic substitutability between the two.
  32
     The prevalence of conditions to produce an evidence or ‘policy’ brieﬁng as a requirement of research funding awards
implies the importance of the medium in the profession’s approach to inﬂuencing public policy. For example, such a re-
quirement is included in most of the UK’s Economic and Research Council grants (https://esrc.ukri.org/funding/guidance-
for-grant-holders/information-for-new-grant-holders/) and International Growth Centre grants (https://www.theigc.org/wp-
content/uploads/2016/12/IGC-Outputs-Guideline.pdf).



                                                               25
of timing etc), this rules out concerns that the letter itself, and not the information within the letter, is
driving the treatment eﬀects of the experiment. Letters and data packages were typically sent two weeks
before the survey from May 2016 onwards. The ﬁrst surveys were conducted in mid-June and lasted until
mid-September.
       The information package was designed to accord exactly with standard communications of the Ethiopian
Public Service. The package, and accompanying letter, was designed in close collaboration with the Min-
istry of Public Service and Human Development, and was sent through the standard bureaucratic channels.
Thus, it was representative of broader service communications, and as far as we understand it would have
been received as such. The accompanying letter stated that, “we have assembled the most up-to-date
administrative data available to provide an understanding of the context your oﬃce works within. We
would like to provide you with this information as an input to improved service delivery within your or-
ganization.” 33 Since the package was shared as a circular, the expectation was for it to be shared widely
among staﬀ within a district oﬃce.
       The data packages make information on core variables of national policy available to a random subset
of Ethiopian oﬃcials in a form they are used to at close to zero cost. The package is a clear reduction in
the marginal cost of aquiring information. Oﬃcials must then decide to read and absorb the information,
and make that decision in the wider incentive environment we have documented above.


4.2      Experimental Results

Table 4 presents the results of the intervention on the errors made by oﬃcials in their claims about
the socio-economic characteristics of the citizenry they serve. The regression framework has the same
structure as (1), the sample is as in previous tables, and the dependent variable is once again a function
of the absolute errors made by oﬃcials. We include a binary treatment variable that takes the value 1 if
the organization received the data package, and 0 if it did not.
       Column 1 of Table 4 presents the unrestricted treatment eﬀect, which does not restrict the comparison
of claims to be within indicator, percentile, district, and so on. We see that it is positive and statistically
insigniﬁcantly diﬀerent from zero at the usual levels. Throughout Table 4, we present p-values of the
treatment eﬀect with two distinct approaches to clustering. In the square and curly brackets below each
treatment coeﬃcient we follow the rest of the paper in clustering at the district government-sector level.
These clusters correspond to the coherent sector teams within the district-government in which many
information distribution choices would most likely be made. However, we note that the treatment is at
  33
    Since the information provided in the circular was a set of statistics about the jurisdictions themselves, they do not suﬀer
from issues of replicability or population validity, a concern with the extrapolation of evidence from other settings (Pritchett
and Sandefur, 2015).



                                                              26
the district level, and therefore we also present the p-value of the treatment coeﬃcient from the same
regression but clustered at the district level. In neither case is the fully unconditional treatment eﬀect in
column 1 signiﬁcant at the usual levels.
   In column 2 of Table 4, we condition on the set of controls used in previous tables, where the treat-
ment eﬀect relates to a comparison within indicator, percentile, sector and region, with individual and
organizational controls. The impact of the data package is now negative and signiﬁcant at the 10% level
(a p-value of 0.08 using our standard clustering approach; a p-value of 0.06 clustering at the level of
treatment). Its magnitude is roughly half of the coeﬃcient on the district government dummy in Table 2,
and large compared to the control variables used for comparisons above. It seems that oﬃcials have taken
the opportunity of the reduction in the cost of information acquisition and increased their knowledge of
the constituencies in which they work.
   Given that we have documented the importance of de facto authority, we control for our core explana-
tory variables from Table 2 in column 3. The treatment eﬀect jumps once we control for these critical
drivers of variation in claim error. The coeﬃcient, -0.28, has a p-value of 0.004 when using our standard
clustering approach and 0.003 when clustering at the treatment level. Once we control for treatment,
the impact of a claim being made by a district oﬃcial increases to -0.44, and the coeﬃcient on manager
continues to be small, which is evidence that the experiment is not driving our results in Tables 2 and
3. A ﬁnal set of controls is the addition of the management indices assessed in Table 3, ‘Management
Practices: Monitoring’ and ‘Management Practices: Other’. The treatment eﬀect continues to be large,
with a coeﬃcient of -0.23, and corresponding p-values of 0.04 with our standard clustering approach and
0.03 when clustering at the treatment level. The management indices both have the expected signs and
are signiﬁcant at the 5% level. Thus, controlling for the key drivers of information acquisition, we see that
a substantial reduction in transaction costs can improve bureaucratic knowledge.
   The treatment eﬀect is robust to using the diﬀerent measurement techniques for aggregating absolute
error. Using proportional rank of error as the dependent variable, shown in column 5, we ﬁnd a treatment
eﬀect of -0.074, with corresponding p-values of 0.03. The eﬀect weakens slightly when we use the skewness-
adjusted z-score, but is still signiﬁcant at the usual levels (column 6). Finally, we note that since the data
package was sent by sectoral oﬃcials at the regional level, we should also test the robustness of the eﬀect
to any correlations between activities within the corresponding regional oﬃce. We therefore cluster at the
region-sector level and the p-value raises to 0.048 (column 7).
   Together, the results support the prediction that substantial reductions in the cost of acquisition for
oﬃcials can improve the accuracy of their information. This could be related to external eﬀorts to improve
the knowledge of public policymakers by presenting key ﬁndings of research in digestible, ‘low-cost’ ways,



                                                     27
and by ‘translating’ research ﬁndings into a medium of communication that reduces the transaction costs
to public servants of absorbing relevant messages. The magnitude of the coeﬃcients in Table 4 are large
relative to the descriptive impacts of incentives documented in Tables 2 and 3, and substantially larger
than many contextual factors, such as making claims over higher poverty districts. Providing a data pack
to an organization improves the accuracy of its oﬃcials’ claims by a similar magnitude to if they were
estimating the conditions of the 5th percentile of poor districts rather than the 90th percentile.
       However, as we have seen in previous tables, the conditions under which public servants work are
highly variable, and estimates relevant to the ‘average’ bureaucrat may not provide a full picture of the
impacts of our treatment. In Table 5, we therefore assess the heterogeneity of the treatment eﬀect along
the critical margins for information acquisition outlined throughout the paper. Column 0 presents our
baseline treatment eﬀect from Table 4 for comparison.
       Column 1 presents the baseline speciﬁcation in column 0 with the addition of an interaction between
the treatment binary and the district government binary. The coeﬃcient is not signiﬁcant at the usual
levels, with a corresponding p-value of 0.14. Similarly, column 2 presents the baseline speciﬁcation with
the addition of an interaction between the manager and treatment indicators. The interaction is once
again not signiﬁcant at the usual levels, and the p-value 0.37. There is thus no evidence that treatment
varied substantially by the level of authority. As expected, in both authority regimes, a large decrease in
the cost of information acquisition increases the equilibrium level of knowledge.34
       Where we should expect to ﬁnd results is in the incentives associated with information. In equilibrium,
agents weigh up the marginal beneﬁt of acquisition with the cost. Where beneﬁts are low, small changes
in the cost of acquisition can have large impacts on the equilibrium calculus for an agent, and our treat-
ment should have large eﬀects. Where the beneﬁts of holding knowledge are high, information holdings
will already be substantial, and thus small changes in costs will lead to limited additional information
absorption. We test these ideas in column 3 of Table 5, where we interact our treatment eﬀect with the
‘Management Practices: Monitoring’ index. The results closely match the theoretical prediction. At the
mean level of management, we observe a similar treatment eﬀect to the baseline. However, there is a large
positive interaction eﬀect, with a coeﬃcient of 0.34, signiﬁcant at the 1% level. As management practices
related to monitoring improve, the mean treatment eﬀect is attenuated. This can be seen clearly in Fig-
ure 5A, where we plot the treatment eﬀect at the deciles of the monitoring management practice index
based on the speciﬁcation of column 3 in Table 5. At the lowest levels of management observed in our
data, the treatment eﬀect is extremely large. The coeﬃcient is -0.87 (p-value 0.00), which is almost three
times the average eﬀect of a claim being made by a district oﬃcial. However, from the 60th percentile of
  34
    Interestingly, there is no evidence of a shift in the distribution of information between managers and employees. This is
consistent with the treatment having been applied consistently within oﬃces.



                                                             28
management practices related to monitoring upwards, the treatment eﬀect is indistinguishable from zero
at the 5% level. The best managed organizations incentivize their oﬃcials such that they already have
suﬃciently precise information that our intervention has no eﬀect.
    In column 4 of Table 5, we interact the treatment binary with the ‘Management Practices: Other’
index. Once again, the interaction is positive and signiﬁcant at the 1% level. The treatment eﬀect at
the mean level of the management index is negative, implying that once again, a higher score on the
management index attenuates our treatment eﬀect towards zero. This can be clearly seen in Figure 5B, in
which we plot the treatment eﬀect at the deciles of the ‘Other’ management index, with all the controls of
the speciﬁcation in column 4 of Table 5. The ﬁgure presents a distribution of coeﬃcients similar to that
for monitoring, with substantial treatment eﬀects at low levels of management, and a null eﬀect at the
60th percentile onwards. Once again, this is consistent with the theory. At low levels of staﬀ alignment,
agents will be more responsive to reductions in costs of information acquisition as there is little crowding
out from free-riding incentives. However, as the organization implements management practices that align
actors’ incentives more closely, the incentive to absorb information lessens and agents aim to free-ride on
each other’s eﬀorts. Overcoming free-riding concerns with a reduction in acquisition costs becomes harder.
    Together, these interactions are consistent with the idea that organizational incentives are important
mediators of the data package treatment. Where incentives for information absorption are strong, oﬃcials
will source information relevant for their core work without external intervention. Where incentives are
weak, there is room for external intervention to improve bureaucratic knowledge. In these poorly managed
environments, external provision of low-cost information can have impacts on the magnitude of oﬃcials’
errors equivalent to or greater than the design of incentives. However, improving the underlying incentives
for information acquisition would yield superior information outcomes without external intervention.


5    Conclusions and Further Research

This paper investigates the scale of errors in the beliefs of public oﬃcials and their determinants across a
large, hierarchical public sector. It measures the subjective beliefs of public oﬃcials as to the characteristics
of the citizens that they serve at all three tiers of government and across ﬁve major sectors. By comparing
these beliefs with benchmark objective data from surveys and administrative data, we identify which
oﬃcials have the most accurate perceptions of the characteristics of their constituents. The paper presents
results from the experimental provision of an evidence brieﬁng that aims to improve the accuracy of public
oﬃcials beliefs. We ﬁnd that such evidence brieﬁngs, now a common tool of economics researchers who
aim to inﬂuence public policy, can improve the accuracy of oﬃcials’ beliefs. However, in line with our
descriptive evidence and consistent with theory, this result is signiﬁcantly mediated by organizational


                                                       29
incentives for information acquisition.
   The distribution of information across bureaucracies is dominantly determined by the incentives of
agents to undertake costly investments in information acquisition and absorption. These considerations
dwarf impacts of the characteristics of the information being considered. Under the appropriate incentives,
oﬃcials can hold relatively accurate beliefs regarding their constituents. Such incentives include holding
de facto decision-making authority, working within an organization whose management practices value
oﬃcials’ holding of information, and being able to access information at very low cost. However, without
these incentives, oﬃcials can make very large errors about the constituents they serve - roughly half of
those bureaucrats we study cannot correctly rank the 25th and 75th percentile of districts in terms of
population in the region in which they specialize. Such errors can have large economic consequences as
they skew the distribution of resources. Relative to an equitable distribution of government resources
across constituents, central oﬃcials would over disburse by 30% to districts at the 75th percentile of the
distribution of errors, and under disburse by 12% to those at the 25th percentile of errors.
   Given the importance of tacit knowledge in bureaucratic settings, the paper presents some of the most
detailed evidence on the nature of information in the public sector we know of. It provides empirical
evidence on the impacts of decentralization on the beliefs of individual public oﬃcials, on the role of man-
agement practices as mediating factors in the impacts of authority, and on the role of external information
provision in a non-market hierarchy. While the empirical setting of this paper is the public sector, it is
reasonable to assume that the same results would be observed where the same hierarchical conditions exist
in the private sector. Our experimental evidence provides quantitative evidence to support the qualitative
impression that the use of evidence in the public sector is mediated by organizational factors. We make
precise the incentive environments in which evidence-brieﬁngs have impacts on bureaucratic beliefs.
   The paper provides a framework for understanding bureaucratic constraints to knowledge diﬀusion in
the public sector. While Brynjolfsson and McElheran (2016) document the rapid adoption of data-driven
decision making in the private sector in the United States, World Bank (2016a) documents the slow rate
of adoption of similar projects in public sectors across the world. Much policy-focused research aims to
translate its ﬁndings into a form digestible by public oﬃcials. However, our results imply that embedding
that knowledge into public organizations is a function of the de facto authority and corresponding incentive
environment in which a public oﬃcial works, as much as it is a function of limiting the costs of absorption
over which the researcher may have some inﬂuence. The open question is how policy-focused researchers
should respond to the incentive environments they face across the public sector.




                                                    30
References


 [1] ABAGISA, J. (2014) “Challenges and Prospects of Balanced scorecard Implementation in the Devel-
     opment Bank of Ethiopia,” Mimeo.

 [2] ACEMOGLU, D., P.AGHION, C.LELARGE, J.VAN REENEN AND F.ZILIBOTI (2007) “Technol-
     ogy, Information, and the Decentralization of the Firm,” Quarterly Journal of Economics 122 (4):
     1759-1799.

 [3] ACEMOGLU, D. AND J. A.ROBNSON (2012) “Why Nations Fail: The Origins of Power, Prosperity
     and Poverty,” New York: Crown Business

 [4] AGHION, P. AND J.TIROLE (1997) “Formal and Real Authority in Organizations,” Journal of
     Political Economy 105 (1): 1–29.

 [5] AGHION, P., N.BLOOM, B.LUCKING, R.SADUN AND J.VAN REENEN (2017) “Turbulence, Firm
     Decentralization and Growth in Bad Times,” National Bureau of Economic Research Working Paper
     No. 23354.

 [6] ANDRADE, P. AND H.LE BIHANA (2013) “Inattentive Professional Forecasters,” Journal of Mon-
     etary Economics 60 (8): 967-982.

 [7] ARGENZIANO, R., S.SEVERINOV AND F.SQUINTANI (2016) “Strategic Information Acquisition
     and Transmission,” American Economic Journal: Microeconomics 8 (3): 119-55.

 [8] AGARWAL, S. AND I.BEN-DAVID (2018) “Loan Prospecting and the Loss of Soft Information,”
     Journal of Financial Economics 129 (3): 608-628.

 [9] AGARWAL, S. AND R.HAUSWALD (2010) “Distance and Private Information in Lending,” Review
     of Financial Studies 23 (7): 2757-2788.

[10] AGARWAL, S. AND R.HAUSWALD (2009) “Authority and Information,” Mimeo.

[11] BANDIERA, O., M.BEST, A.KHAN AND A.PRAT (2018) “Incentives and the Allocation of Author-
     ity in Organizations: A Field Experiment with Bureaucrats,” Mimeo: London School of Economics.

[12] BANERJEE, A., R.HANNA, J.KYLE, B. A.OLKEN AND S.SUMARTO (2018) “Tangible Infor-
     mation and Citizen Empowerment: Identiﬁcation Cards and Food Subsidy Programs in Indonesia,”
     Journal of Political Economy 126 (2): 451-491.

[13] BARDHAN, P. (2002) “Decentralization of Governance and Development,” Journal of Economic
     Perspectives 16 (4): 185-205.

[14] BARDHAN, P. AND D.MOOKHERJEE (2006). “Decentralisation and Accountability in Infrastruc-
     ture Delivery in Developing Countries,” Economic Journal, 116 (508): 101-127.

[15] BENDER, S., N.BLOOM, D.CARD, J.VAN REENEN AND S.WOLTER (2018) “Management Prac-
     tices, Workforce Selection, and Productivity,” Journal of Labor Economics 36 (S1): S371-S409.

[16] BERTRAND, M., R.BURGESS, A.CHAWLA AND G.XU (2016) “The Costs of Bureaucratic Rigidity:
     Evidence from the Indian Administrative Service,” University of Chicago Working Paper.

[17] BESLEY, T. AND T. PERSSON (2011) “Pillars of Prosperity: The Political Economics of Develop-
     ment Clusters,” Princeton: Princeton University Press.


                                                 31
[18] BLOOM, N. AND J.VAN REENEN (2007) “Measuring and Explaining Management Practices Across
     Firms and Countries,” Quarterly Journal of Economics 122 (4): 1351-1408.

[19] BLOOM, N., B.EIFERT, A.MAHAJAN, D.MCKENZIE AND J.ROBERTS (2013) “Does Manage-
     ment Matter? Evidence from India,” Quarterly Journal of Economics 128 (1): 1–51.

[20] BLOOM, N., R.SADUN AND J.VAN REENEN (2012) “The Organization of Firms Across Countries,”
     Quarterly Journal of Economics 127 (4): 1663-1705.

[21] BLOOM, N., C.PROPPER, S.SEILER AND J.VAN REENEN (2015) “The Impact of Competition
     on Management Practices in Public Hospitals,” Review of Economic Studies 82 (2): 457–489.

[22] BROWN, M., W.SCHALLER, S.WESTERFIELD AND M.HEUSLER (2012) “Information or Insur-
     ance? On the Role of Discretion in Relationship Lending,” University of St. Gallen Working Paper.

[23] BRYNJOLFSSON, E. AND K.MCELHERAN (2016) “The Rapid Adoption of Data-Driven Decision-
     Making,” American Economic Review 106 (5): 133-39.

[24] CAMPBELL COLLABORATION (2019) “The Campbell Collaboration Online Library,” Available at
     https://campbellcollaboration.org/library.html (Accessed 4th February 2019)

[25] CALLEN, M., S.GULZAR, A.HASANAIN, M.Y.KHAN AND A.REZAEE (2018) “Data and Policy
     Decisions: Experimental Evidence from Pakistan,” Stanford University Working Paper.

[26] CRAWFORD, G. AND C.HARTMANN (2008) “Decentralisation in Africa: A Pathway Out of Poverty
     and Conﬂict?” Amsterdam: Amsterdam University Press.

[27] CRAWFORD, V. P. AND J.SOBEL (1982). “Strategic Information Transmission,” Econometrica:
     Journal of the Econometric Society : 1431-1451.

[28] CREMER, J., A.ESTACHE AND P.SEABRIGHT (1994) “The Decentralization of Public Services:
     Lessons from the Theory of the Firm,” World Bank Policy Research Working Paper 1345

[29] D’AGOSTINO, R.B., A.J.BELANGER AND R.B.D’AGOSTINO JR (1990) “A Suggestion for Using
     Powerful and Informative Tests of Normality,” American Statistician 44 (4): 316-321.

[30] DAL BÓ, E., F.FINAN AND M.A.ROSSI (2013) “Strengthening State Capabilities: The Role of
     Financial Incentives in the Call to Public Service,” Quarterly Journal of Economics 128 (3): 1169-
     1218.

[31] DAL BÓ, E., F.FINAN, N.Y.LI AND L.SCHECHTER (2018) “Government Decentralization Under
     Changing State Capacity: Experimental Evidence from Paraguay,” National Bureau of Economic
     Research Working Paper No. 24879.

[32] DESSEIN, W. (2002) “Authority and Communication in Organizations,” Review of Economic Studies
     69 (4): 811-838.

[33] DHALIWAL, I. AND R.HANNA (2017) “The Devil is in the Details: The Successes and Limitations
     of Bureaucratic Reform in India,” Journal of Development Economics 124: 1-21.

[34] DI PEI, H. (2015) “Communication with Endogenous Information Acquisition,” Journal of Economic
     Theory 160: 132-149.

[35] DUFLO, E. (2017) “Richard T. Ely Lecture: The Economist as Plumber,” American Economic Review
     107 (5): 1-26.



                                                  32
[36] DUGGAN, M. AND F.M.SCOTT MORTON (2006) “The Distortionary Eﬀects of Government Pro-
     curement: Evidence from Medicaid Prescription Drug Purchasing,” Quarterly Journal of Economics
     121 (1): 1-30.

[37] EHRBECK, T. AND R.WALDMANN (1996) “Why are Professional Forecasters Biased? Agency
     Versus Behavioral Explanations,” Quarterly Journal of Economics 111 (1): 21-40.

[38] ELLIOTT, H. AND J.POPAY (2000) “How Are Policy Makers Using Evidence? Models of Research
     Utilisation and Local NHS Policy Making,” Journal of Epidemiology & Community Health 54 (6):
     461-468.

[39] FINAN, F., B.OLKEN AND R.PANDE (2017) “The Personnel Economics of the Developing State,”
     Handbook of Economic Field Experiments 2: 467-514

[40] FISHER, J.R.B., J.MONTAMBAULT, K.P.BURFORD, T.GOPALAKRISHNA, Y.J.MASUDA,
     S.M.W.REDDY, K.TORPHY AND A.I.SALCEDO (2018) “Knowledge Diﬀusion Within a Large Con-
     servation Organization and Beyond,” PLoS ONE 13(3).

[41] GABAIX, X., D.LAIBSON, G.MOLOCHE AND S.WEINBERG (2006) “Costly Information Acqui-
     sition: Experimental Analysis of a Boundedly Rational Model,” American Economic Review 96 (4):
     1043-1068

[42] GANT, J., C.ICHNIOWSKI AND K.SHAW (2002) “Social Capital and Organizational Change in
     High-Involvement and Traditional Work Organizations,” Journal of Economics & Management Strat-
     egy 11 (2): 289-328.

[43] GARICANO, L. (2000) “Hierarchies and the Organization of Knowledge in Production,” Journal of
     Political Economy 108 (5): 874-904.

[44] HANNA, R., S.MULLAINATHAN AND J.SCHWARTZSTEIN (2014) “Learning Through Noticing:
     Theory and Evidence from a Field Experiment,” Quarterly Journal of Economics 129 (3): 1311–1353.

[45] HARRIS, R.D.F. AND E.TZAVALIS (1999) “Inference for Unit Roots in Dynamic Panels Where the
     Time Dimension is Fixed,” Journal of Econometrics 91 (2): 201–226

[46] HASTINGS, J.S. AND J.M.WEINSTEIN (2008) “Information, School Choice, and Academic Achieve-
     ment: Evidence from Two Experiments,” Quarterly Journal of Economics 123 (4): 1373-1414.

[47] HATTORI, M., K.SHINTANI AND H.UCHIDA (2015) “The Repository of Soft Information Within
     Bank Organizations,” Journal of Money, Credit and Banking 47 (4): 737-770.

[48] IM, K.S., M.H.PESARAN AND Y.SHIN (2003) “Testing for Unit Roots in Heterogeneous Panels,”
     Journal of Econometrics 115 (1): 53–74.

[49] JESSOE, K. AND D.RAPSON (2014) “Knowledge Is (Less) Power: Experimental Evidence from
     Residential Energy Use,” American Economic Review 104 (4): 1417-38.

[50] LANDIER, A., V.B.NAIR, AND J.WULF (2007). “Trade-Oﬀs in Staying Close: Corporate Decision
     Making and Geographic Dispersion,” Review of Financial Studies 22 (3): 1119-1148.

[51] LIBERTI, J.M. (2018) “Initiative, Incentives, and Soft Information,” Management Science 64 (8):
     3714-3734.

[52] MASAKI, T., S.CUSTER, A.ESKENAZI, A.STERN AND R.LATOURELL (2017) “Decoding Data
     Use: How Do Leaders Source Data and Use it to Accelerate Development?” Williamsburg, VA:
     AidData at the College of William & Mary.

                                                 33
[53] MOOKHERJEE, D. (2006) “Decentralization, Hierarchies, and Incentives: A Mechanism Design
     Perspective,” Journal of Economic Literature 44 (2): 367-390.

[54] MURALIDHARAN, K. AND V.SUNDARARAMAN (2011) “Teacher Performance Pay: Experimental
     Evidence from India,” Journal of Political Economy 119 (1): 39-77.

[55] MURALIDHARAN, K., P.NIEHAUS AND S.SUKHTANKAR (2016) “Building State Capacity: Ev-
     idence from Biometric Smartcards in India,” American Economic Review 106 (10): 2895-2929.

[56] MYATT.D, P. AND C.WALLACE (2012) “Endogenous Information Acquisition in Coordination
     Games,” Review of Economic Studies 79 (1): 340–374

[57] MYERSON, R.B. (1981) “Optimal Auction Design,” Mathematics of Operations Research 6 (1): 58-73.

[58] OATES, W.E. (1999) “An Essay on Fiscal Federalism,” Journal of Economic Literature 37 (3): 1120-
     1149.

[59] OATES, W.E. (2005) “Toward A Second-Generation Theory of Fiscal Federalism,” International Tax
     and Public Finance 12 (4): 349–373.

[60] OLKEN, B.A. (2007) “Monitoring Corruption: Evidence from a Field Experiment in Indonesia,”
     Journal of Political Economy 115 (2): 200-249.

[61] PAWSON, R. (2006) “Evidence-Based Policy: A Realist Perspective,” London: Sage Publications.

[62] PARKHURST, J. (2017) “The Politics of Evidence: From Evidence-Based Policy to the Good Gov-
     ernance of Evidence,” Routledge Studies in Governance and Public Policy, Abingdon: Routledge.

[63] PEPINSKY, T.B., J.H.PIERSKALLA AND A.SACKS (2017) “Bureaucracy and Service Delivery,”
     Annual Review of Political Science 20: 249-268.

[64] PETERSEN, M.A. (2004) “Information: Hard and Soft,” Mimeo.

[65] POMERANZ, D. (2015) “No Taxation Without Information: Deterrence and Self-Enforcement in the
     Value Added Tax,” American Economic Review 105 (8): 2539-69.

[66] PRITCHETT, L. AND J.SANDEFUR (2015) “Learning from Experiments When Context Matters,”
     American Economic Review 105 (5): 471-75.

[67] QIAN, J., P.E.STRAHAN AND Z.YANG (2015) “The Impact of Incentives and Communication
     Costs on Information Production and Use: Evidence from Bank Lending,” Journal of Finance 70 (4):
     1457-1493.

[68] RASUL, I. AND D.ROGGER (2018) “Management of Bureaucrats and Public Service Delivery: Ev-
     idence from the Nigerian Civil Service,” Economic Journal 128: 413-446.

[69] RASUL, I., D.ROGGER AND M.WILLIAMS (2018) “Management and Bureaucratic Eﬀectiveness:
     Evidence from the Ghanaian Civil Service,” Mimeo, University College London.

[70] ROYSTON, P. (1992) “Comment on sg3. 4 and an Improved D’Agostino Test,” Stata Technical Bul-
     letin 1(3).

[71] SMITH, A.F.M. (1996) “Mad Cows and Ecstasy: Chance and Choice in an Evidence-Based Society,”
     Journal of the Royal Statistical Society 159 (3): 367-383.

[72] SOMANI, R. (2018) “Information and Public Service Delivery: Evidence from an E-Governance
     Initiative,” Mimeo, University College London.

                                                 34
[73] STEIN, J.C. (2002) “Information Production and Capital Allocation: Decentralized Versus Hierar-
     chical Firms,” Journal of Finance 57 (5): 1891-1921.

[74] STIGLITZ, J.E. (2017) “The Revolution of Information Economics: The Past and the Future,” Na-
     tional Bureau of Economic Research Working Paper No. 23780.

[75] TEREDA, A. (2014) ”Design and Implementation of Balanced Scorecard in the Development Bank
     of Ethiopia,” Mimeo.

[76] UCHIDA, H., G.F.UDELL AND N.YAMORI (2012) “Loan oﬃcers and relationship lending to SMEs,”
     Journal of Financial Intermediation 21 (1): 97-122.

[77] WORLD BANK (2012) “Better Results from Public Sector Institutions,” Washington, DC: World
     Bank

[78] WORLD BANK (2016a) “World Development Report 2016: Digital Dividends,” Washington, DC:
     World Bank

[79] WORLD BANK (2016b) “Ethiopia Public Expenditure Review,” Washington, DC: World Bank


Data Appendix

Measuring the Errors of Oﬃcialdom

This section describes in detail the creation of the error variables, their relative scale, and a descriptive
assessment of their basic determinants. For each individual surveyed, we obtained their subjective as-
sessments (‘claims’) of demographic characteristics of their constituency, of the sector-speciﬁc variables
relating to their constituency, and of the variables relating to the public service rules.
   Where individuals stated that they did not know the answer at all, a value of 0 was entered to signal
that the respondent did not know the answer to a particular question at all. Of the 891 individuals who
make up our common-claims sample used in the core analysis, 275 (31%) of individuals respond ‘Don’t
Know’ to at least one of the questions asked. 181 individuals (20%) respond ‘Don’t Know’ to all of the
questions in the information module. There is, therefore, selection into the sample of respondents who
provide non-zero estimates. Regressing individual characteristics on a binary variable that takes the value
1 if a respondent answers ‘Don’t Know’ to all questions, we ﬁnd evidence that district oﬃcials are less
likely to do so. We make two comments regarding the selection into the error sample: ﬁrstly, it seems a
reasonable assumption that the probability of responding ‘Don’t Know’ is positively associated with the
most extreme (latent) errors, and hence our estimates of the diﬀerence across tiers can be interpreted as a
lower bound. Secondly, the error data provide consistent estimates of the accuracy of information that is
present within a public oﬃce in Ethiopia – for example, if one entered a public oﬃce at random in Ethiopia
and asked about a particular service delivery indicator, the information that one would receive would be
reﬂected within the error data we use for our regressions. To ensure that these 0 claims do not aﬀect the


                                                      35
analysis on the errors, a missing value is inserted in place of the 0 for the error variables. Furthermore, a
‘Don’t Know’ binary indicator is created that takes the value 1 whenever a respondent does not know the
answer to a question.
   Our benchmark data is regarded as the best available in the Ethiopian context. One proxy of the
quality of the benchmarking data is their stability over time. By assessing the extent to which indicators
vary within districts over data collection rounds, we aim to understand whether the data reﬂect a noisy
collection process. For example, if the ratio of pupils per school has a high variance from one year to the
next within the same district, this could be seen as a warning sign as to the quality of measurement. Table
A2 presents summary statistics and tests for the presence of a random walk for each of the indicators used
in the analysis where panel data are available. Column 4 shows the estimated autoregressive parameter
from a Dickey Fuller Regression, controlling for district ﬁxed eﬀects and column 5 shows the estimated
autoregressive parameter from a Dickey Fuller Regression, controlling for district ﬁxed eﬀects and district-
speciﬁc linear time trends. The estimated coeﬃcients are less than 1, which indicate stability in the trend,
for all indicators except for three: the amount of land used for agricultural purposes; the amount of land
used for pastoral purposes; and the number of business licenses issued. In addition, column 6 reports
the z-statistics and column 7 the associated p-value from a unit-root test based on Harris and Tzavalis
(1999) where there are suﬃcient data to do so. Column 8 reports the ﬁxed-N critical value at the 5%
level and column 9 the ﬁxed-N test statistic from Im, Pesaran and Shin (2003), allowing for linear time
trends. These tests are appropriate for panel data settings with short timeframes. The results of the tests
provide signiﬁcant evidence against the existence of a unit-root process, and therefore a random walk, for
the tested series. These results provide support to the notion that the benchmarking data are consistent
over time.
   For each statistic we study, we subtract the corresponding benchmark data sourced from survey or
administrative data from the claim of the respondent. For questions related to the rules of the service,
which are uniform across all respondents regardless of sector and tier, the claim is subtracted from the
benchmark data and the absolute value generated. For the demographic and sector claims, the format
of these diﬀer slightly by tier. District oﬃcials were asked directly about their district, and the relevant
benchmark indicators subtracted from these responses. The regional oﬃcials were asked about the 25th,
50th and 75th percentile districts of individual variables separately. The federal oﬃcials were asked, “Which
region, out of the following options, do you work on most/think that you know best/feel most comfortable
answering questions about?” They were then asked about the 25th, 50th and 75th percentile districts of
individual variables within that region. This ensures that each oﬃcial is responding to questions on the
districts that they are most familiar with, and work most closely on. Federal oﬃcials across sectors chose



                                                     36
the full distribution of regions, and the distribution of choices closely matches population and district
numbers.
   The raw errors vary signiﬁcantly, as documented in the main text. The distribution of the errors
suggests that civil servants, on average, tend to overestimate the value of the indicator, with the average
error of all but three of the indicators we study being positive (the three indicators with a negative actual
error are: the proportion of rural inhabitants; the primary pupil-class ratio; and the hectares of land used
for pastoral purposes). Since it is diﬃcult to compare errors across diﬀerent indicators, the overall skewness
coeﬃcient for the z-score of the actual error is 0.4. A test of skewness relative to the normal distribution
is rejected at the usual levels. Individually, the indicators tend to be positively skewed (have distributions
with longer right tails). Six of the 20 indicators have a negative skewness coeﬃcent: proportion of rural
inhabitants (-3.02); the antenatal care rate (-0.49); the rate of births delivered by a skilled attendant
(-0.22); the share of households living on subsistence agriculture (-0.73); the proportion of income from
regional block grants (0.64); and the proportion of income from own sources (-0.56).
   In the core analysis, we aim to compare the claims of oﬃcials at multiple tiers of government relating to
the same district. We therefore focus on districts for which we have subjective assessments from each tier
of government. This requires a district to be one of the 25th/50th/75th percentiles for at least one of the
indicators we study, since these are the districts over which federal and regional oﬃcials made assessments.
We therefore drop roughly a third (24) of the 66 districts on which we collected data to ensure that each
claim has counterparts at all three tiers of government.
   We create a data set where the unit of observation is a claim of a bureaucrat about a district.
As described, each federal and regional oﬃcial is asked about three districts for each indicator (the
25th/50th/75th percentiles of that indicator), and so each bureaucrat-indicator-percentile is an obser-
vation for these oﬃcials. The claim of the district oﬃcial about a speciﬁc indicator in her district is a
single observation. Each response is associated with the corresponding district census ID.
   In a very small number of cases, there seems to have been a data entry issue, where additional digits
were included or excluded in the recording of responses to non-fractional variables. This leads us to some
errors that are oﬀ by a very signiﬁcant factor (of 94 at the 95th percentile of errors). These are not
reﬂective of the wider variation in the data. Two-thirds of absolute errors are less than 1 in proportional
terms. The most extreme values seem outside the realm of feasiblity for even the least informed bureaucrat.
Since a plausible explanation is that these outliers reﬂect data entry issues rather than outsized errors,
we therefore censor the top 5% of raw errors for each of the variables. The interpretation of the paper is
therefore analysis of the lowest 95% of errors that public oﬃcials make, rather than of the top 5%.
   Absolute values are then taken of all errors. At this stage, we have a data set that is not comparable



                                                     37
across, but only within indicators. Making an error of 0.5 in the percentage of rural inhabitants is large for
that indicator, but tiny for the population indicator. We therefore undertake a series of transformations
of the absolute errors to make them comparable across indicators. We take three approaches.
   Our core measure is created by constructing z-scores that are the value of the absolute error minus
the mean value across the distribution of absolute errors within the indicator, all divided by the standard
deviation of absolute errors across the distribution of absolute errors within the indicator. The interpre-
tation of this z-score is a unitless measure of how many standard deviations away from the benchmark
data an oﬃcial’s assessment is. Comparisons across indicators therefore take into account the empirical
distribution of errors for each indicator. They compare the scale of an oﬃcial’s error relative to the errors
made by all other oﬃcials in our data set.
   One issue with the above measure is that it does not take into account potential skewness of the
underlying indicators. Table A2 provides information on the Pearson moment coeﬃcient of skewness for
each of the indicators. It is evident that the non-fraction indicators exhibit right-skew, with positive
coeﬃcients larger than 1 in all cases. For non-fraction indicators, a test for skewness relative to the
normal distribution is conducted based on D’Agostino et al. (1990) and Royston (1992). The test statistic
is approximately normal under the null hypothesis and allows for standard hypothesis tests. Thus, if
the p-value associated with the test statistic is less than or equal to 0.05, we classiﬁed the variable as
signiﬁcantly skewed and replace the indicator with its log, consistent with Duggan and Morton (2004).
For those that are not signiﬁcantly skewed, no transformation is applied. The ‘skewness-adjusted’ z-score
is calculated on this set of values.
   A third approach is to use the ordinal rank of the absolute error within the entire distribution of
absolute errors within the indicator (again consistent with Duggan and Morton, 2004). We rank all errors
according to size, give each a corresponding rank, and divide these ranks by the maximum rank within
the distribution of absolute errors within the indicator. The rank-proportion scores are therefore free of
distributional assumptions. They can be interpreted as the percentile of error made by an oﬃcial relative
to other oﬃcials in our data set.


Measuring the incentive environment

To construct our management indices, we followed the best-practices laid out in the World Management
Survey literature (such as in Bloom and Van Reenen, 2007; Bloom et al., 2012). Management interviews
were ‘double blind’ so: (i) managers were not told in advance they were being scored or shown a score grid;
(ii) enumerators were given no information on the organization’s performance. Only managers (directors)
were enumerated the World Management Survey and the measures of organizational management practices



                                                     38
we use for our core analysis average management scores over the most senior divisional-bureaucrat reports
where there are multiple directors.
   The scores on each question are converted into normalized z-scores (so are continuous variables with
mean zero and variance one by construction) and are increasing in the commonly understood notion of
‘better management’. For example, we assume greater monitoring corresponds to better management prac-
tice. We then take unweighted means of the underlying z-scores to aggregate the questions on monitoring
shown in Table A3 and separately all other management practices.
   We complement our measure of alignment with a second measure by asking civil servants directly
about the extent of alignment they feel to the organization and their daily activities. In the civil servants
survey, we asked oﬃcials, “To what extent do you believe that the organization’s mission is aligned to your
own mission?” and “How consistent with your mission are the various tasks and activities assigned to you
on a day-to-day basis?” To measure the extent to which the organization as a whole and the individual
tasks assigned to the civil servant are consistent with her own mission, we create an alignment index that
is the mean of the organization average response to the former of these questions and the proportion of
employees in the organization that state “I am contributing to fulﬁlling that mission on an everyday basis”
to the latter.


Basic Determinants of Bureaucrat Errors

What are the descriptive determinants of these errors? Table A4 undertakes complementary ANOVA
and OLS analyses where the dependent variable is the z-score of errors across variables outlined above.
We focus on variables related to the hierarchical structure of the bureaucracy, and the marginal costs
and beneﬁts of information acquisition. Column 1 of Table A4 presents partial sum of squares for these
variables with corresponding F-statistic p-values presented in parentheses. Column 2 provides coeﬃcients
for the corresponding variables within a formal regression framework for consistency with our later tables.
   The ﬁrst panel tests whether the hierarchical structure of government matters for the quality of oﬃcials’
information along three margins. We see that tier of government is of substantial importance, and has
a higher partial sum of squares than any of the other ‘structure’ variables. Its p-value is 0.00, which is
consistent with the notion that the structure of formal authority is a critical determinant of information
absorption. In column 2, where the coeﬃcients are from a regression of the variables listed in Table A4 on
the z-score of errors, we see that the district government binary has a p-value of 0.02, previewing future
results.
   We can also assess these unconditional results graphically. Figure A3 presents the average errors of
centralized and district oﬃcials in their claims about each of the 42 districts on which we focus our core



                                                     39
analysis. The set of claims over which these averages are made are the same as in our regressions such
that they are balanced across variables and percentiles. The scatterplot is ranked by the average error of
claims made about each district, and we can immediately see that some districts are fundamentally more
diﬃcult to make assessments about. Corresponding to the results in Table 2, we see that the unconditional
means of claims by district oﬃcials are typically below those of oﬃcials in centralized organizations across
the distribution. In 75% of cases, the district oﬃcials’ assessments are more accurate than those of the
corresponding central government oﬃcials.
   We also test the extent to which being of managerial rank impacts the scale of an oﬃcial’s errors, and
ﬁnd that the sum of squares is very low and the associated p-value is 0.83. The corresponding conditional
coeﬃcient in column 2 is 0.01, with a p-value of 0.86. We also ﬁnd little evidence that there are substantial
diﬀerences in errors across sectors, consistent with the idea that it is the structure of incentives, rather
than the topic of focus, that matters for information acquisition.
   Though not shown in the table, we have also investigated the impact of environmental features on the
scale of errors made. These are lower in districts with a lower poverty rate, smaller population, and those
that are less remote. However, these features explain only 1.3% of the variation we observe, and once
again highlight the primacy of incentives in determining the quality of information a public oﬃcial has.
As indicated in Figure 3, the mean error varies substantially across claimed about districts. Therefore,
there must be unobserved aspects of individual districts that make information costlier to acquire.
   It seems natural to hypothesize that agents would have a greater information advantage under de facto
agent-authority in environments where information is costlier to acquire. However, increases in costs to
both agents typically have an ambiguous eﬀect on the relative distribution of information. For example,
diﬀerences in the shape of the cost function at the speciﬁc levels of investment will play a role in determining
the eﬀect on the relative distribution. It is therefore an empirical question as to whether the diﬀerential
between principal and agent increases as the contracting environment becomes more diﬃcult. To assess
this in our setting, Table A6 presents the results of a quantile regression of our preferred speciﬁcation over
the quantiles of absolute error. The coeﬃcient on the district dummy is negative and signiﬁcant at the
1% level across all percentiles of absolute error, implying that district oﬃcials’ claims are closer to the
benchmark data than centralized oﬃcials’ claims at all levels of absolute error. There is no clear pattern
across the percentiles in terms of the size of coeﬃcients, and 6 of the 9 are between -0.20 and -0.27. Thus,
the impact of formal authority in this setting is not mediated by the scale of the error to which it relates,
but rather has broadly positive impacts on information acquisition.
   The marginal costs of information acquisition may be determined by features of the variables we test
oﬃcials on, the characteristics of the responding oﬃcial, their organization’s infrastructure for dissem-



                                                      40
inating information, and features of the local environment. The second panel of Table A4 outlines a
corresponding series of variables that increase or decrease the marginal costs of information acquisition.
First we focus on features of the claims, and see that the speciﬁc indicator on which oﬃcials make claims
explains a signiﬁcant portion of the variation in overall error rates, as one might expect. The percentile of
an indicator (whether it is the 25th, 50th or 75th percentile district in the data) similarly impacts the scale
of error, with the average being a more familiar quantity to oﬃcials than other points in the distribution.
These results imply the importance of controlling for indicator features in our core regressions.
   Being a more experienced or educated bureaucrat may lower the cost of information acquisition. How-
ever, the coeﬃcients in column 1 corresponding to bureaucratic characteristics indicate that individual
demographics have little impact on the oﬃcials’ overall likelihood of making errors. The corresponding
regression coeﬃcients are all insigniﬁcant at the usual levels. The sum of squares is higher for organiza-
tional features, consistent with theory. For example, Aghion and Tirole (1997) argues that “it is always
optimal for the ﬁrm to be in a situation of overload” where managers credibly commit to lower monitoring
of employees through increased span of control. A higher span of control therefore leads to lower errors
on average among oﬃcials, as employees increase their investment in information. We see that the corre-
sponding coeﬃcient on span of control in column 2 of Table A4 is negative, implying that operating under
a higher span of control leads to oﬃcials making lower errors in their claims.
   To ensure consistency with the main analysis, we include district ﬁxed eﬀects in Table A4. To explore
the notion that simple proximity to local constituents (such as being an oﬃcial at a local-level organization)
is driving diﬀerences in information quality, we can drop these ﬁxed eﬀects and include measures of
‘remoteness’ of the district: the average travel time of constituents to the nearest urban area, the average
elevation of the district, the average slope of the district, and the road density of the district. The partial
sum of squares of all of these variables combined is 2.76. While non-trivial (and accounted for in our main
speciﬁcations with district government ﬁxed eﬀects), these variables combined account for less variation in
errors than variables capturing management practices or decentralization. In the OLS regression similar
to column 2 of table A4, removing district ﬁxed eﬀects and including these variables, none are statistically
signiﬁcant at usual levels and a joint Wald test of their signiﬁcance gives a p-value of 0.546. These results
are consistent with the argument put forward in Cremer et al. (1994) that simple proximity is unlikely to
be a key driver in diﬀerences in information across tiers of government.
   The ﬁnal panel of Table A4 assesses factors related to the marginal beneﬁts of information acquisition.
We split our measures of management practices into those related to monitoring and all others. Manage-
ment practices related to monitoring explain the largest portion of the variance in errors. The coeﬃcient
on monitoring-related management practices in column 2 of Table A4 is negative, as we would expect, and



                                                      41
signiﬁcant at the 5% level.
   Alignment between the mission of an organization and its employees is a feature of the principal-
agent relationship frequently discussed in the literature. Table A4 includes our complementary measures
of alignment. The measure based on the management practices approach explains substantially more
variation than that based on direct questioning. However, the coeﬃcients on neither variable are signiﬁcant
at the usual levels in column 2 with p-values of 0.138 and 0.431 respectively.


Robustness of Core Results

To assess the robustness of the core descriptive results to our methods of measurement, we present a
series of robustness checks in Table A5. The ﬁrst column of Table A5 re-runs our preferred regression,
but disaggregating the central tier of government binary into the federal and regional tiers, with region as
the excluded category. The coeﬃcient on district government is -0.31 with a p-value of 0.00, equivalent to
that in our preferred regression, and the federal government binary has a value of 0.04 and is insigniﬁcant
at the usual levels, implying that there is no statistical diﬀerence between the errors made by federal
and regional oﬃcials. As described above, the federal and regional organization structure is equivalent
and functionally distinct to the district. In columns 2 to 5 of Table A5, we return to a binary district
variable and vary the level at which we cluster the standard errors in our preferred regression. Taking the
conservative perspective that information regarding a speciﬁc district may be freely shared among oﬃcials
across government, we cluster at the ‘claimed about’ district in column 2. The district coeﬃcient continues
to be signiﬁcant at the 1% level. Similarly, we may be concerned that errors are correlated across the
respondent’s claims, within their organization, or across the region they work in through their regional
sector oﬃces (who provide them with regulations and technical assistance). We therefore cluster at the
respondent level, the organization-of-respondent level, and the region-sector level in columns 3, 4 and 5 of
Table A5 respectively. In each case, the coeﬃcient on the district government binary is signiﬁcant at the
1% level. We can also rerun our preferred speciﬁcation using proportional rank as an alternative measure
of the magnitude of errors. Column 6 of Table A5 presents the result, which implies a district oﬃcial makes
an error that is 10 percentage points lower than a central bureaucrat. Returning to comparison with the
poverty rate, this magnitude is equivalent to 37 times the magnitude of the coeﬃcient on the poverty rate
using this dependent variable, and thus akin to making claims over the 25th percentile of poor districts
rather than the 62nd. As outlined in section 2.4, we may be concerned about using absolute errors in the
presence of skewness of the underlying indicators. We therefore re-run our preferred speciﬁcation on the
‘skewness-adjusted z-score’ and ﬁnd that the coeﬃcient on district government is even larger in magnitude
with a coeﬃcient of -0.39, and a p-value of 0.00. Distinct measurement approaches produce qualitatively



                                                    42
similar results in this setting.


Public Service Rules

Tacit knowledge of local conditions can be characterized as ‘soft information’, which cannot be veriﬁed by
parties other than those who collected it. Contract theory aligned with Myerson (1981) and summarized in
Mookherjee (2006) has pointed out that if information is ‘hard’ then under the Revelation Principle there
is no reason to expect formal delegation to lead to superior information. Since under credible commitment
and without constraints on contract complexity and information exchange, the central organization can
perfectly mimic the delegated organization. As Petersen (2004) argues, the crucial attribute of models
of information in hierarchy is that information assets are subjective in nature, such that contracts must
be incomplete. Much of the recent contract literature has focused on the diﬀerential implications of hard
versus soft information, especially in ﬁnancial intermediation and credit decisions (Stein, 2002).
   It could be argued that the Ethiopian government’s ‘Public Service Rules’, written in hardcopy for all
oﬃcials, can be characterized as ‘hard information’. Given its veriﬁability, Dessein (2002) argues that soft
information leads to ineﬃcient centralized decision making stemming from noisy communication, while
communication is preferable to delegation if there is little uncertainty so that headquarters could make
decisions primarily based on hard or public information. Table A7 presents a regression in which the
dependent variable is the z-score of the absolute error of claims about the public service rules. We ﬁnd
that across our speciﬁcations the Federal Government makes lower errors over hard information than either
of the other two tiers of government. Column 7 is equivalent to our preferred speciﬁcation in Table 2, and
it implies Federal oﬃcials make errors over the public service rules that are 0.09 of a standard deviation
lower than regional colleagues, who make errors that are statistically insigniﬁcantly diﬀerent to those at
the district level, at the usual levels of signiﬁcance. This result is robust to clustering at the respondent
level (column 8 of Table A7) and the region-sector level (column 9), and to using the skewness adjusted z-
score (column 11). Column 10 indicates that the diﬀerence is equivalent to a 4 percentage point reduction
in errors.




                                                     43
Table 1: Characteristics of Civil Servants
Means and standard deviations
                                                                                 (1) All                  (2) Federal              (3) Regional                (4) Local

Organizational Characteristics

Number of organizations                                                           382                           5                        54                       323
Number of bureaucrats per organization                                            5.71                        56.60                     8.67                      4.43
Number of managers per organization                                               1.22                        6.80                      2.57                      0.90
Number of employees per organization                                              3.58                        49.20                     5.41                      2.56
Span of control (employees per manager)                                           7.79                        27.28                    15.98                      5.97

Official's Characteristics
Number of bureaucrats                                                             1831                         280                      431                      1120
Age                                                                              35.30                        35.87                    38.10                     34.07
                                                                                 [8.85]                      [9.10]                    [8.74]                    [8.58]
Years in position                                                                 2.70                        2.29                      2.98                      2.69
                                                                                 [2.74]                      [2.88]                    [2.87]                    [2.64]
Years in organization                                                             7.34                        5.80                      7.97                      7.48
                                                                                 [7.13]                      [6.47]                    [7.27]                    [7.18]
Years in civil service                                                           13.13                       12.87                     15.49                     12.29
                                                                                 [8.88]                      [8.98]                    [9.00]                    [8.65]
Number of service organizations worked in                                         2.78                        3.02                      3.08                      2.55
                                                                                 [2.01]                      [2.07]                    [2.40]                    [1.73]
Grade                                                                             7.93                        7.53                     10.89                      6.90
                                                                                 [13.06]                     [9.92]                   [17.97]                   [11.21]
Education [undergraduate degree=1]                                                 0.82                       0.96                      0.90                      0.76
                                                                                 [0.38]                      [0.19]                    [0.30]                    [0.43]
Education [masters degree=1]                                                      0.11                        0.35                      0.22                      0.01
                                                                                 [0.32]                      [0.48]                    [0.42]                    [0.12]
Gender [female=1]                                                                 0.20                        0.27                      0.23                      0.18
                                                                                 [0.40]                      [0.44]                    [0.42]                    [0.38]
Manager                                                                           0.25                        0.12                      0.32                      0.26
                                                                                 [0.44]                      [0.33]                    [0.47]                    [0.44]
Notes: Standard deviations are in parentheses. The unit of observation for the first panel is the civil servant and for the second panel the organization. Only employees and
managers responded to the information module and hence only these individuals are used for the statistics in the first panel. These statistics are based on the survey sample
and not on administrative data. The span of control is taken from the survey question "How many personnel do you manage?"; the average is taken for an organization if
there is more than one manager; and the organization-level statistics are reported (each organization has an equal weight in the `By Organization' statistics). `Manager' refers
to managers within the organization. Grade is the official civil service grade of the professional civil servant, ranging from 1 to 17 in the sample. The 2013-14 National Civil
Service Human Resource Statistics Abstract (Ministry of Civil Service) records the overall female employee percentage to be 35%; this statistic includes frontline staff. Figures
are rounded to two decimal places.
Figure 1A: All Respondents' Population Estimates and Actual Population
                   Population and claims across all respondents


                                                                                                                                   20
   25



   20                                                                                                                              15




                                                                                                                        Percent
   15
                                                                                                                                   10

   10


                                                                                                                                    5
      5                                                                              Actual
                                                                                     Estimated
                                                                                     Median population in sample
      0
                                                                                                                                    0
             0          100000       200000       300000      400000      500000                                                    -400000              -200000                    0                 200000                  400000
                                           Population                                                                                                                              Error
             Only claims below 500,000 shown                                                                                            Only errors for claims below 500,000 shown


	Notes:	Figure	1A	shows	the	histograms	of	actual	populations	(black)	and	population	claims	(blue	dashed)	for	all	claims	on	the	left-hand	panel;		the	right-hand	panel	show	histograms	of	actual	errors	for	all	claims.		Actual	errors	
are	the	claim	of	the	respondent	minus	the	actual	population	value.		The	black	dashed	line	represents	the	median	actual	population	in	the	sample	of	districts.		All	claims	refers	to	claims	over	the	75th	percentile,	50th	percentile,	
and	25th	percentile	district	in	terms	of	population	within	the	(chosen)	region	for	the	federal	and	regional	respondents;	and	claims	over	own	districts	for	district	level	respondents.		The	sample	includes	all	employees	and	
managers	in	the	Ethiopian	Civil	Servants	Survey	that	were	asked	the	information	module.		In	both	panels,	observations	corresponding	to	populaton	claims	over	500,000	are	not	included	for	presentation	purposes.	



Figure 1B: Federal and Regional Respondent Population Estimates and Actual Population
                                             Population error for 25th percentile district                                                                     Population claim for 25th percentile district
                                                                                                                                                                as a percentage of the actual population
             25
                                                                                                                                   25



             20
                                                                                                                                   20



             15
   Percent




                                                                                                                                   15
                                                                                                                        Percent




             10                                                                                                                    10



              5                                                                                                                     5




              0                                                                                                                     0
              -200000            -100000              0         100000        200000            300000        400000                     0      100     200        300     400      500       600    700       800      900      1000
                                                                Error                                                                                         Population claim over actual population (%)
                  Only errors for claims below 500,000 shown                                                                            Only errors for claims below 500,000 and up to 10 times the value of the data shown



                                             Population error for 50th percentile district                                                                     Population claim for 50th percentile district
             25
                                                                                                                                                                as a percentage of the actual population
                                                                                                                                   25


             20
                                                                                                                                   20



             15
   Percent




                                                                                                                                   15
                                                                                                                        Percent




             10                                                                                                                    10



              5                                                                                                                     5



              0                                                                                                                     0
              -200000        -100000              0           100000        200000           300000      400000                          0      100     200        300     400      500       600    700       800      900      1000
                                                                Error                                                                                         Population claim over actual population (%)
                  Only errors for claims below 500,000 shown                                                                            Only errors for claims below 500,000 and up to 10 times the value of the data shown



                                             Population error for 75th percentile district                                                                     Population claim for 75th percentile district
             25                                                                                                                                                 as a percentage of the actual population
                                                                                                                                   25


             20
                                                                                                                                   20


             15
   Percent




                                                                                                                                   15
                                                                                                                         Percent




             10
                                                                                                                                   10



              5                                                                                                                     5



              0                                                                                                                     0
              -200000            -100000              0         100000        200000            300000        400000                     0       100     200        300      400        500    600     700       800       900      1000
                                                                Error                                                                                          Population claim over actual population (%)
                  Only errors for claims below 500,000 shown                                                                            Only errors for claims below 500,000 and up to 10 times the value of the data shown

Notes: Figure 1B left-hand panels show histograms of actual errors for the 25th percentile district, 50th percentile district, and 75th percentile district. Actual errors are the claim of the respondent minus the actual population
value. The black dashed line represents the median actual population in the sample of districts. The sample is all federal and regional employees and managers in the Ethiopian Civil Servants Survey that were asked the
information module. The right-hand panel shows histograms of the population claim divided by the actual population of the district multiplied by 100 to convey the size of the population claim relative to the data in percentage
terms. The solid gray line at 100 represents an accurate claim. In both panels, observations corresponding to populaton claims over 500,000 are not included for presentation purposes. In the right-hand panel, observations
where the claim is more than 10 times (1000% of) the actual data are not shown.
Figure 1C: Distribution of Biased Weights
All individuals in federal and regional governments


                                                                Individual population weighting errors for 50th percentile woreda
                                                            5

       Federal and regional respndent effective weighting

                                                            4




                                                            3




                                                            2




                                                            1


                                                                                                                                                                      Respondent weighting
                                                                                                                                                                      Equity per capita
                                                            0
                                                                  0               100          200                 300                              400
                                                                                        Rank of weighting

Managers in federal and regional governments


                                                                Managers' population weighting errors for 50th percentile woreda
                                                            5
       Federal and regional respndent effective weighting




                                                            4




                                                            3




                                                            2




                                                            1


                                                                                                                                                                      Respondent weighting
                                                                                                                                                                      Equity per capita
                                                            0
                                                                 0           20          40          60                    80                       100
                                                                                        Rank of weighting


Organization-level average population weighting error, federal and regional governments


     Organisation average population weighting errors for 50th percentile woreda
                                                            5
       Federal and regional respndent effective weighting




                                                            4




                                                            3




                                                            2




                                                            1


                                                                                                                                                                           Respondent weighting
                                                                                                                                                                           Equity per capita
                                                            0
                                                                 0                10          20                 30                              40
                                                                                        Rank of weighting


Notes: The figure shows the organization-average estimated population weights for the median district in the region in terms of population. The top panel shows the observations for all employees and
managers in the sample; the middle panel shows the observations for all managers in the sample; the bottom panel shows the organization-level averages of all employees and managers in the sample. The
weights are calcuated based on respondent estimates of the population across the 25th, 50th, and 75th percentile district in the region. The black line represents equity per capita -- the correct weights if the
actual data were used; the black circles represent the weight for the median disrict based on the region and federal civil servant respondents' estimates of the populations of the 25th, 50th, and 75th percentile
districts. The organization average weight is then used for the purpose of this figure. The sample for the top panel is all federal and regional employees and managers that responded to the information module
in the Ethiopian Civil Servants Survey; the sample for the middle panel is all federal and regional managers that responded to the information module in the Ethiopian Civil Servants Survey; the sample for the
bottom panel is all federal and regional employees and managers that responded to the information module in the Ethiopian Civil Servants Survey.
Figure 2: Distributions of Errors Across Tiers of Government
Figure 2A: CDF of Errors Using Z-Scores

                              Cumulative Distribution Functions of Z-Scores of Errors
                    1




                .8




                .6
     Density




                .4




                .2
                                                                                                                           Federal
                                                                                                                           Regional
                                                                                                                           District Government
                    0
                         -2            0              2                  4                  6                  8
                                                    Z-score of error



Figure 2B: CDF of Errors Using Skewness Adjusted Z-Scores

               Cumulative Distribution Functions of Skew-Adjusted Z-Scores of Errors
               1




               .8




               .6
     Density




               .4




               .2
                                                                                                                               Federal
                                                                                                                               Regional
                                                                                                                               District Government
               0
                        -6             -4              -2                     0                 2                   4
                                            Skew-Adjusted Z-Scores of Error


Figure 2C: CDF of Errors Using Ranks

                               Cumulative Distribution Functions of Ranks of Errors
                    1




                .8




                .6
     Density




                .4




                .2
                                                                                                                           Federal
                                                                                                                           Regional
                                                                                                                           District Government
                    0
                         0             .2             .4                 .6                .8                  1
                                              Rank of error (proportion)


Notes: The figure shows the cumulative distribution functions of the federal, regional, and district government level respondent's errors across all policy
domains. The top panel shows the errors in terms of z-score within the indicator. The middle panel shows the error in terms of skewness-adjusted z-
score within the indicator; the skewness-adjusted z-score is the standard z-score in cases of variables with a 0 to 1 support and for variables that are not
significantly skewed according to the test from D'Agostino, Balanger, and D'Agostino (1990); for those variables that are not 0 to 1 in support and are
significantly skewed, the log of the variable is taken and the z-score of this is the skewness-adjusted z-score. The bottom panel shows the errors in
terms of rank within the indicator as a proportion of the maximum rank. The sample used in the analysis is the set of respondents making claims about
the same district. The black solid line refers to federal claims; the gray dashed line to regional claims; and the blue dotted line to district government
claims.
Table 2: Hierarchy and Information
Dependent Variable: A binary indicator of whether respondent doesn't know in columns (1) and (2); Z-score of errors across all policy domains in columns (3) to (10)
OLS Estimates
Standard Errors: Clustered at the district government-sector level

                                                                        (1) Don't know           (2) Don't know        (3) Unconditional (4) Unconditional (5) Unconditional                                                 (7) District                                                             (10) Individual
                                                                                                                                                                                                    (6) Sector FEs                        (8) Indicator FEs (9) Percentile FEs
                                                                        Unconditional              Conditional           Decentralized        Manager         Hierarchies                                                  government FEs                                                             Characteristics

District Government [Yes=1]                                                   -0.17                    -0.17                    -0.27                                            -0.27                    -0.28                    -0.30                    -0.31                    -0.31                   -0.33
                                                                             [0.034]                  [0.033]                  [0.084]                                          [0.084]                  [0.084]                  [0.086]                  [0.087]                  [0.087]                 [0.087]
Manager [Yes=1]                                                               0.011                   0.0040                                            0.011                    0.023                    0.020                    0.014                    0.016                    0.015                   0.034
                                                                             [0.031]                  [0.029]                                          [0.066]                  [0.066]                  [0.067]                  [0.065]                  [0.066]                  [0.066]                 [0.069]
Sector fixed effects                                                          No                      Yes                       No                       No                       No                       Yes                     Yes                      Yes                      Yes                      Yes
District government fixed effects                                             No                      Yes                       No                       No                       No                       No                      Yes                      Yes                      Yes                      Yes
Indicator fixed effects                                                       No                      Yes                       No                       No                       No                       No                      No                       Yes                      Yes                      Yes
Percentile fixed effects                                                      No                      Yes                       No                       No                       No                       No                      No                       No                       Yes                      Yes
Individual controls                                                           No                      Yes                       No                       No                       No                       No                      No                       No                       No                       Yes
Adjusted R-squared                                                           0.022                    0.13                     0.012                  -0.00066                   0.012                   0.0099                   0.059                    0.067                    0.072                    0.069
Observations [clusters]                                                    2144 [98]                2144 [98]                1455 [92]                1455 [92]                1455 [92]                1455 [92]               1455 [92]                1455 [92]                1455 [92]                1455 [92]
Central government respondents [percentage]                                594 [67%]                594 [67%]                429 [62%]               429 [62%]                429 [62%]                429 [62%]                429 [62%]                429 [62%]                429 [62%]               429 [62%]
District government respondents [percentage]                               297 [33%]                297 [33%]                258 [38%]               258 [38%]                258 [38%]                258 [38%]                258 [38%]                258 [38%]                258 [38%]               258 [38%]
Notes: Clustered standard errors are in parentheses. Standard errors clustered at the district government-sector interacted level. All columns report OLS estimates. The unit of analysis in all columns is the claim. The sample used for the analysis is the set of respondents that make claims about the
same districts. The dependent variable is a binary indicator for whether the respondent could not provide a claim in columns (1) and (2); the z-score of absolute error across all policy domain items in columns (3) to (12). Sector fixed effects are indicators for whether the claim is within an agriculture,
education, health, revenue, or trade organization. Percentile fixed effects are indicators for the 75th, 50th or 25th percentile for the respective indicator for the respective region. District government fixed effects identify the district that the respondent is making claims about. Indicator fixed effects control
for the service delivery indicator within which the claim is made. Manager is equal to one if the claim is from a director. Individual characteristics included in the specification: tenure in the civil service of the respondent making the claim; tenure in the position; tenure in the organization; an indicator for
whether the respondent is female; an indicator for whether the respondent holds an undergraduate degree; an indicator for whether the respondent holds a masters degree. Figures are rounded to two significant figures.
Table 3: Information, Incentives and Alignment
Dependent Variable: Z-score of errors across all policy domains in all columns
OLS Estimates
Standard Errors: Clustered at the district government-sector level
                                                                                                                                                                            Incentives

                                                                                   (1) Management practices                     (2) Noise controls                        (3) Combined                 (4) Interaction Decentralized              (5) Interaction Manager

District Government [Yes=1]                                                                                                                                                     -0.27                                  -0.34                                  -0.27
                                                                                                                                                                                [0.16]                                 [0.16]                                 [0.16]
Manager [Yes=1]                                                                                  0.011                                 -0.026                                  -0.029                                 -0.021                                 -0.035
                                                                                                [0.070]                                [0.076]                                 [0.076]                                [0.077]                                [0.076]
Management Practice: Monitoring                                                                  -0.17                                  -0.21                                   -0.20                                  -0.14                                  -0.17
                                                                                                [0.067]                                [0.067]                                 [0.067]                                [0.080]                                [0.067]
Management Practice: Other                                                                       0.28                                    0.31                                    0.32                                   0.32                                   0.31
                                                                                                [0.097]                                 [0.11]                                  [0.11]                                 [0.10]                                 [0.11]
District Government x Management Practice: Monitoring                                                                                                                                                                  -0.19
                                                                                                                                                                                                                       [0.14]
Manager x Management Practice: Monitoring                                                                                                                                                                                                                     -0.10
                                                                                                                                                                                                                                                             [0.090]
Sector, indicator, percentile fixed effects                                                      Yes                                     Yes                                    Yes                                    Yes                                     Yes
District government fixed effects                                                                No                                      No                                     No                                     No                                      No
Region fixed effects                                                                             Yes                                     Yes                                    Yes                                    Yes                                     Yes
Individual controls                                                                              Yes                                     Yes                                    Yes                                    Yes                                     Yes
Marginal costs and noise controls                                                                No                                      Yes                                    Yes                                    Yes                                     Yes
Adjusted R-squared                                                                              0.034                                   0.061                                  0.063                                  0.064                                   0.063
Observations [clusters]                                                                       1455 [92]                               1455 [92]                              1455 [92]                              1455 [92]                               1455 [92]
Central government respondents [percentage]                                                   429 [62%]                              429 [62%]                               429 [62%]                              429 [62%]                              429 [62%]
District government respondents [percentage]                                                  258 [38%]                              258 [38%]                               258 [38%]                              258 [38%]                              258 [38%]
Notes: Clustered standard errors are in parentheses. Standard errors clustered at the district government-sector interacted level, allowing for error correlation within regions, sectors, and districts.All columns report OLS estimates.The unit of analysis in all columns is
the claim.The tiers of government are grouped into District Government and Centralized Government which incorporates both Federal and Regional tiers, hence the omitted category for tier in the regressions is Centralized Government.The sample used for the
analysis is the set of respondents that make claims about the same districts.The dependent variable is the z-score of absolute errors across all policy domains.Sector fixed effects are indicators for whether the claim is within an agriculture, education, health, revenue,
or trade organization.Percentile fixed effects are indicators for the 75th, 50th or 25th percentile for the respective indicator for the respective region.District government fixed effects identify the district that the respondent is making claims about.Indicator fixed effects
control for the service delivery indicator within which the claim is made.Management practice: Monitoring is the organization-average z-score for the monitoring practice of management as per Bloom and Van Reenen (2007); Management practice: Other is the mean
of the organization average z-scores for targeting, incentives, roles, flexibility, staffing, and staff involvement. Manager is equal to one if the claim is from a manager.Individual characteristics included in the specification:tenure in the civil service of the respondent
making the claim;tenure in the position; tenure in the organization;an indicator for whether the respondent is female; an indicator for whether the respondent holds an undergraduate degree; an indicator for whether the respondent holds a masters degree; and an
indicator for whether the respondent is a manager. Marginal costs include the proportion of managers in the organization that state that a management information system is in place; the organization average response to "In what proportion of projects would you say
information flows effectively between individuals in the same organization?"; the organization average response to "In what proportion of projects would you say Information flows effectively between organizations?"; the organization average response of managers to
"How many personnel do you manage?"; the organization average response of employees to "How many people would you say regularly give you tasks as part of your formal work duties?";alignment (the z-score of the row mean of the organization-average
response to "To what extent do you believe that your organization's mission is aligned to your own mission?" and the organization proportion of selecting "I am contributing to fulfilling that mission on an everyday basis" to "How consistent with your mission are the
various tasks and activities assigned to you on a day-to-day basis?").Noise controls include the time of day of the survey, day-of-survey fixed effects, enumerator fixed effects, an indicator for the enumerator's subjective assessment of the quality of the interview; an
indicator for the duration of the interview (decile of duration fixed effects); controls for the average tenure of the managers in the organization and the proportion of female managers in the organization.Figures are rounded to two significant figures.
Figure 3: Incentives, Decentralization and Information
                                        The Effect of Decentralization on Information
                                       by Centiles of Management practice: Monitoring
                                  .5
    Z-Score of Absolute Error




                                  0




                                 -.5




                                 -1


                                                                                           Coefﬁcient District Government [Yes=1]
                                                                                           Conﬁdence interval (95%)
                                -1.5
                                        0       20      40       60      80      100
                                       Percentile of Management practice: Monitoring


Notes: The figure shows the OLS coefficient on District Government [Yes=1] for each decile of Management practice: Monitoring, in a
regression also conditional on Management practice: Other; Individual controls; Sector fixed effects; Indicator fixed effects, Percentile
fixed effects, Marginal cost controls and Noise controls. 95% confidence intervals are show in the gray lines, with standard errors
clustered at the district government-sector interacted level.
                         Table 4: Information Experiment

                         Dependent Variable: Z-score of errors across policy domains in columns (1) to (4) and (7); rank of error (proportion) in column (5); and skewness-adjusted z-score in column (6)

                         OLS estimates in all columns
                         Standard Errors: Clustered at the district government-sector level in columns (1) to (6) and at the region-sector level in column (7)


                                                                                                                                                    (2) Conditional                (3) Baseline with               (4) Management                   (5) Rank Error            (6) Skewness-Adjusted              (7) Clustering at
                                                                                                                  (1) Treatment only
                                                                                                                                                   Treatment Effect                    treatment                       practices                     (Proportion)                  Z-Score Error               Region-Sector Level


                         Treated District [Yes=1]                                                                          0.053                          -0.15                           -0.28                            -0.23                          -0.074                           -0.2                           -0.23
                                                                                                                          [0.087]                        [0.085]                         [0.098]                           [0.11]                        [0.032]                          [0.11]                          [0.11]
                           P-value of coefficient clustered at district government-sector level                            {0.54}                        {0.078}                        {0.0048}                         {0.036}                         {0.022}                        {0.086}                         {0.048}
                           P-value of coefficient clustered at district level                                              (0.51)                        (0.057)                        (0.0031)                         (0.033)                         (0.024)                        (0.053)                         (0.033)
                         District Government [Yes=1]                                                                                                                                      -0.44                            -0.41                           -0.11                          -0.39                           -0.41
                                                                                                                                                                                          [0.17]                           [0.17]                        [0.048]                          [0.17]                          [0.16]
                         Manager [Yes=1]                                                                                                                                                 -0.021                           -0.026                          -0.011                        -0.0074                          -0.026
                                                                                                                                                                                         [0.077]                         [0.076]                         [0.020]                         [0.066]                        [0.080]
                         Management Practice: Monitoring                                                                                                                                                                   -0.18                          -0.018                         -0.027                           -0.18
                                                                                                                                                                                                                         [0.069]                         [0.024]                         [0.088]                        [0.064]
                         Management Practice: Other                                                                                                                                                                         0.28                           0.038                          0.090                            0.28
                                                                                                                                                                                                                           [0.11]                        [0.038]                          [0.13]                          [0.14]
                         Sector, region, percentile, indicator fixed effects                                               No                                                                                                              Yes
                         Individual controls                                                                               No                                                                                                              Yes
                         Marginal costs and noise controls                                                                 No                                                                                                              Yes
                         Adjusted R-squared                                                                             -0.00018                          0.057                           0.062                          0.066                           0.092                           0.10                            0.066
                         Number of observations [clusters]                                                              1455 [92]                       1455 [92]                       1455 [92]                      1455 [92]                       1455 [92]                       1455 [92]                       1455 [38]
                        Notes: All columns report OLS regression coefficients and standard errors in brckets. Standard errors in brackets. Standard errors clustered at the district government-sector interacted level in columns (1) to (6) and at the region-sector interacted level in column (7). P-values are reported in
                        curly brackets. P-values associated with clustering at the district level are reported in round brackets. The unit of analysis in all columns is the claim. The sample used for the analysis is the set of respondents that make claims about the same districts. The dependent variable is the z-score
                        error
spective region. Indicator     foreffects
                           fixed   any claim
                                          are made
                                              indicator
                                                    within  any policy
                                                        variables      domain
                                                                  for which specific
                                                                              in columns
                                                                                     variable
                                                                                           (1)the claim
                                                                                               to (4)   iscolumn
                                                                                                      and  being made
                                                                                                                  (7). The
                                                                                                                       over.
                                                                                                                           dependent variable
                                                                                                                             Management       is theMonitoring
                                                                                                                                          practice: rank of theis
                                                                                                                                                                absolute
                                                                                                                                                                  the organization-average
                                                                                                                                                                          error in terms of the proportion
                                                                                                                                                                                              z-score       ofmonitoring
                                                                                                                                                                                                      for the the total number
                                                                                                                                                                                                                         practice
                                                                                                                                                                                                                                of of
                                                                                                                                                                                                                                   claims for the indicator
                                                                                                                                                                                                                                      management            in column
                                                                                                                                                                                                                                                    as per Bloom   and(5).
                                                                                                                                                                                                                                                                       VanThe
                                                                                                                                                                                                                                                                           Reenen
                                                                                                                                                                                                                                                                              dependent variable
                                                                                                                                                                                                                                                                                  (2007);        is the skewness-adjusted
                                                                                                                                                                                                                                                                                          Management    practice: Other is the
                                                                                                                                                                                                                                                                                                                             z-score
                                                                                                                                                                                                                                                                                                                               mean in
                                                                                                                                                                                                                                                                                                                                     of column (6). The tiers
                                                                                                                                                                                                                                                                                                                                        the organization      of gove
                                                                                                                                                                                                                                                                                                                                                         average   z-s
                        figures.
Table 5: Heterogeneous Treatment Effects
Dependent Variable: Z-score of errors across policy domains
OLS estimates in all columns
Standard Errors: Clustered at the district government-sector level


                                                                                                                                                                                                            (4) Interaction: other
                                                                                   (0) Baseline treatment (1) Interaction: district                 (2) Interaction:             (3) Interaction:
                                                                                                                                                                                                                 management
                                                                                           effect               government                              manager                    monitoring
                                                                                                                                                                                                                   practices


Treated District [Yes=1]                                                                      -0.23                        -0.025                         -0.20                         -0.26                         -0.34
                                                                                              [0.11]                        [0.19]                        [0.12]                        [0.10]                        [0.11]
District Government [Yes=1]                                                                   -0.41                        -0.089                         -0.41                         -0.46                         -0.43
                                                                                              [0.17]                        [0.28]                        [0.17]                        [0.16]                        [0.16]
Manager                                                                                      -0.026                        -0.032                         0.068                        -0.018                        -0.015
                                                                                             [0.076]                       [0.075]                        [0.13]                       [0.076]                       [0.076]
Management Practice: Monitoring                                                               -0.18                         -0.18                         -0.18                         -0.43                         -0.18
                                                                                             [0.069]                       [0.073]                       [0.069]                        [0.11]                       [0.065]
Management Practice: Other                                                                     0.28                          0.30                          0.28                          0.30                         0.023
                                                                                              [0.11]                        [0.12]                        [0.11]                        [0.11]                        [0.14]
Treated District x District Government [Yes=1]                                                                              -0.37
                                                                                                                            [0.25]
Treated District x Manager                                                                                                                                -0.12
                                                                                                                                                          [0.13]
Treated District x Management Practice: Monitoring                                                                                                                                       0.34
                                                                                                                                                                                        [0.13]
Treated District x Management Practice: Other                                                                                                                                                                         0.38
                                                                                                                                                                                                                     [0.14]
Sector, region, percentile, indicator fixed effects                                                                                                       Yes
Individual controls                                                                                                                                       Yes
Marginal costs and noise controls                                                                                                                         Yes
Adjusted R-squared                                                                           0.066                         0.067                         0.065                         0.071                         0.071
Number of observations [clusters]                                                          1455 [92]                     1455 [92]                     1455 [92]                     1455 [92]                     1455 [92]
Notes: All columns report OLS regression coefficients and standard errors in brckets. Standard errors clustered at the district government-sector interacted level. The unit of analysis in all columns is the claim. The sample used for the analysis is the set
projects would you say information flows effectively between individuals in the same organization?"; the organization average response to "In what proportion of projects would you say Information flows effectively between
organizations?"; the organization average response of managers to "How many personnel do you manage?"; the organization average response of employees to "How many people would you say regularly give you tasks as part
of your formal work duties?"; alignment (the z-score of the row mean of the organization-average response to "To what extent do you believe that your organization's mission is aligned to your own mission?" and the organization
proportion of selecting "I am contributing to fulfilling that mission on an everyday basis" to "How consistent with your mission are the various tasks and activities assigned to you on a day-to-day basis?"). Noise controls include
the time of day of the survey, day-of-survey fixed effects, enumerator fixed effects, an indicator for the enumerator's subjective assessment of the quality of the interview; an indicator for the duration of the interview (decile of
duration fixed effects); controls for the average tenure of the managers in the organization and the proportion of female managers in the organization. Figures are rounded to two decimal places or two significant figures.
Figure 4: Heterogeneous Treatment Effects by Management practices
4A: Treatment and Management practice: Monitoring


                                                The Effect of Treatment on Information
                                            by Centiles of Management practice: Monitoring
                                   1




                                  .5
      Z-Score of Absolute Error




                                   0




                                  -.5




                                  -1
                                                                                                                                             Coefﬁcient 1[Treated District]
                                                                                                                                             Conﬁdence interval (95%)

                                        0          20          40                  60                   80                 100
                                            Percentile of Management practice: Monitoring


4B: Treatment and Management practice: Other


                                                The Effect of Treatment on Information
                                              by Centiles of Management practice: Other
                                   1




                                  .5
      Z-Score of Absolute Error




                                   0




                                  -.5




                                  -1                                                                                                         Coefﬁcient 1[Treated District]
                                                                                                                                             Conﬁdence interval (95%)

                                        0          20          40                  60                   80                 100
                                              Percentile of Management practice: Other

Notes: The figure shows the effect of the treatment on the z-score of errors acorss all policy domains by different percentles of Management practice: Monitoring in the top figure; and by different
percentiles of Management practice: Other in the bottom figure. The sample used for the analysis is the set of respondents that make claims about the same districts. Management practices and
errors are in z-scores. Treated district is a binary indicator equal to one if the district received an information package Regressions are conditional on Management practices; Individual controls;
Sector fixed effects; Indicator fixed effects, Percentile fixed effects, Marginal cost controls and noise controls, Region fixed effects; and an indicator for whether the claim is from a District
Government. 95% confidence intervals are shown in the gray lines, with standard errors clustered at the district government-sector level. Management practice: Monitoring is the organization
average of the monitoring topic, construced from the z-scores of each individual item noted in table A3. Management practice: Other is the organization average of the average across the z-scores
of targeting, incentives, roles, flexibility, staffing, and staff involvement, detailed in table A3.
TA1: List of Surveyed Organizations

Tier of Governance        Region                                             Organization Name

     Federal                   -                   Federal Ministries of Agriculture; Education; Health; Revenue; and, Trade
     Regional           Addis Ababa          Addis Ababa City Administration Bureaus of Education; Health; Revenue; and, Trade
     Regional                Afar               Afar Regional Bureaus of Agriculture; Education; Health; Revenue; and, Trade
     Regional              Amhara             Amhara Regional Bureaus of Agriculture; Education; Health; Revenue; and, Trade
     Regional        Benishangul Gumuz Benishangul Gumuz Regional Bureaus of Agriculture; Education; Health; Revenue; and, Trade
     Regional            Dire Dawa      Dire Dawa City Administration Bureaus of Agriculture; Education; Health; Revenue; and, Trade
     Regional             Gambella           Gambella Regional Bureaus of Agriculture; Education; Health; Revenue; and, Trade
     Regional               Harar              Harar Regional Bureaus of Agriculture; Education; Health; Revenue; and, Trade
     Regional              Oromia             Oromia Regional Bureaus of Agriculture; Education; Health; Revenue; and, Trade
     Regional              SNNPR              SNNPR Regional Bureaus of Agriculture; Education; Health; Revenue; and, Trade
     Regional              Somali             Somali Regional Bureaus of Agriculture; Education; Health; Revenue; and, Trade
     Regional               Tigray             Tigray Regional Bureaus of Agriculture; Education; Health; Revenue; and, Trade
      District               Afar              Afar Awash Fentale Agriculture, Education, Health, Revenue, and Trade Offices
      District               Afar                   Afar Telalak Agriculture, Education, Health, Revenue, and Trade Offices
      District               Afar                    Afar Teru Agriculture, Education, Health, Revenue, and Trade Offices
      District             Amhara                 Amhara Awabel Agriculture, Education, Health, Revenue, and Trade Offices
      District             Amhara            Amhara Basona Worana Agriculture, Education, Health, Revenue, and Trade Offices
      District             Amhara      Amhara Borena (Former Debresina) Agriculture, Education, Health, Revenue, and Trade Offices
      District             Amhara              Amhara Chefa Gula Agriculture, Education, Health, Revenue, and Trade Offices
      District             Amhara                 Amhara Dejen Agriculture, Education, Health, Revenue, and Trade Offices
      District             Amhara            Amhara Enarj Enawaga Agriculture, Education, Health, Revenue, and Trade Offices
      District             Amhara                 Amhara Gidane Agriculture, Education, Health, Revenue, and Trade Offices
      District             Amhara              Amhara Jabitahnan Agriculture, Education, Health, Revenue, and Trade Offices
      District             Amhara              Amhara Jile Timuga Agriculture, Education, Health, Revenue, and Trade Offices
      District             Amhara                Amhara Kutaber Agriculture, Education, Health, Revenue, and Trade Offices
      District             Amhara                Amhara Simada Agriculture, Education, Health, Revenue, and Trade Offices
      District       Benishangul Gumuz     Benishangul Gumuz Dibate Agriculture, Education, Health, Revenue, and Trade Offices
      District       Benishangul Gumuz      Benishangul Gumuz Yasso Agriculture, Education, Health, Revenue, and Trade Offices
      District            Gambella          Gambella Gambella Zuria Agriculture, Education, Health, Revenue, and Trade Offices
      District            Gambella               Gambella Abobo Agriculture, Education, Health, Revenue, and Trade Offices
      District             Oromia                   Oromia Ale Agriculture, Education, Health, Revenue, and Trade Offices
      District             Oromia                 Oromia Amigna Agriculture, Education, Health, Revenue, and Trade Offices
                                         Oromia Arsi Negelle District government Agriculture, Education, Health, Revenue, and Trade
      District             Oromia
                                                                                      Offices
      District             Oromia                 Oromia Babile Agriculture, Education, Health, Revenue, and Trade Offices
      District             Oromia               Oromia Bako Tibe Agriculture, Education, Health, Revenue, and Trade Offices
      District             Oromia                  Oromia Begi Agriculture, Education, Health, Revenue, and Trade Offices
      District             Oromia                Oromia Dedessa Agriculture, Education, Health, Revenue, and Trade Offices
      District             Oromia              Oromia Digluna Tijo Agriculture, Education, Health, Revenue, and Trade Offices
      District             Oromia       Oromia Gida Ayana (Gida Kiremu) Agriculture, Education, Health, Revenue, and Trade Offices
      District             Oromia           Oromia Goro Gutu (Goro) Agriculture, Education, Health, Revenue, and Trade Offices
      District             Oromia                 Oromia Guduru Agriculture, Education, Health, Revenue, and Trade Offices
      District             Oromia               Oromia Haro Maya Agriculture, Education, Health, Revenue, and Trade Offices
      District             Oromia                 Oromia Hitosa Agriculture, Education, Health, Revenue, and Trade Offices
      District             Oromia             Oromia Jardega Jarte Agriculture, Education, Health, Revenue, and Trade Offices
      District             Oromia                  Oromia Jeldu Agriculture, Education, Health, Revenue, and Trade Offices
      District             Oromia                 Oromia Kofale Agriculture, Education, Health, Revenue, and Trade Offices
      District             Oromia                 Oromia Mesela Agriculture, Education, Health, Revenue, and Trade Offices
      District             Oromia              Oromia Midaga Tola Agriculture, Education, Health, Revenue, and Trade Offices
      District             Oromia                  Oromia Nono Agriculture, Education, Health, Revenue, and Trade Offices
      District             Oromia                  Oromia Seru Agriculture, Education, Health, Revenue, and Trade Offices
      District             Oromia                  Oromia Siraro Agriculture, Education, Health, Revenue, and Trade Offices
      District             Oromia             Oromia Tikur Enchini Agriculture, Education, Health, Revenue, and Trade Offices
      District             Oromia                 Oromia Wadera Agriculture, Education, Health, Revenue, and Trade Offices
      District             Oromia              Oromia Were Jarso Agriculture, Education, Health, Revenue, and Trade Offices
      District             SNNPR                  SNNPR Amaro Agriculture, Education, Health, Revenue, and Trade Offices
      District             SNNPR                 SNNPR Analimo Agriculture, Education, Health, Revenue, and Trade Offices
      District             SNNPR                 SNNPR Basketo Agriculture, Education, Health, Revenue, and Trade Offices
      District             SNNPR              SNNPR Benatsemay Agriculture, Education, Health, Revenue, and Trade Offices
      District             SNNPR                   SNNPR Bona Agriculture, Education, Health, Revenue, and Trade Offices
      District             SNNPR                  SNNPR Chere Agriculture, Education, Health, Revenue, and Trade Offices
      District             SNNPR                   SNNPR Dale Agriculture, Education, Health, Revenue, and Trade Offices
      District             SNNPR                  SNNPR Decha Agriculture, Education, Health, Revenue, and Trade Offices
      District             SNNPR               SNNPR Doyo Gena Agriculture, Education, Health, Revenue, and Trade Offices
      District             SNNPR                SNNPR Gomibora Agriculture, Education, Health, Revenue, and Trade Offices
      District             SNNPR             SNNPR Hawassa Zuriya Agriculture, Education, Health, Revenue, and Trade Offices
      District             SNNPR                  SNNPR Kucha Agriculture, Education, Health, Revenue, and Trade Offices
      District             SNNPR                SNNPR Shebedino Agriculture, Education, Health, Revenue, and Trade Offices
      District             SNNPR                 SNNPR Wenago Agriculture, Education, Health, Revenue, and Trade Offices
      District             Somali                  Somali Afdem Agriculture, Education, Health, Revenue, and Trade Offices
      District             Somali        Somali Erer District government Agriculture, Education, Health, Revenue, and Trade Offices
      District             Somali                 Somali Harshin Agriculture, Education, Health, Revenue, and Trade Offices
                                         Somali Jijiga Zuria District government Agriculture, Education, Health, Revenue, and Trade
      District             Somali
                                                                                      Offices
      District             Somali              Somali Kebri Beyah Agriculture, Education, Health, Revenue, and Trade Offices
      District              Tigray                  Tigray Erob Agriculture, Education, Health, Revenue, and Trade Offices
      District              Tigray             Tigray Gulo Mekeda Agriculture, Education, Health, Revenue, and Trade Offices
      District              Tigray            Tigray Hintalo Wajerat Agriculture, Education, Health, Revenue, and Trade Offices
      District              Tigray             Tigray Tahtay Koraro Agriculture, Education, Health, Revenue, and Trade Offices
      District              Tigray               Tigray Wereilehi Agriculture, Education, Health, Revenue, and Trade Offices
Table A2: Description of Survey and Administrative Data
Summary statistics
                                                                                                                                                                                                                                                                                                                                                                       (8)
                                                                                                                                                                                                                                                                                                                            (6)                   (7)
                                                                                                                                                                                                                                                                   (4) Autogregressive        (5) Autogregressive                                            Fixed-N exact critical
                                                                                                                                                                                                                                                                                                                    Z Statistic from     P-value from Harris                                 (9)                                                 (10)
                                                                                                                                                                                                                 (2) Standard                                       parameter from DF          parameter from DF                                                value for the Im-
                                                                                                                                                        Source                          (1) Mean                                            (3) Skewness                                                           Harris Tsavalis Unit  Tsavalis Unit Root                         Fixed-N Im-Pesaran-                                           N
                                                                                                                                                                                                                   deviation                                       regression, no time         regression, linear                                            Pesaran-Shin test at
                                                                                                                                                                                                                                                                                                                  Root Test (linear time  Test (linear time                          Shin Test statistic
                                                                                                                                                                                                                                                                          trends                   time trends                                                the 5% level (linear
                                                                                                                                                                                                                                                                                                                         trends)               trends)
                                                                                                                                                                                                                                                                                                                                                                  time trends)
Demographic and Socio-economic Characteristics
What do you think is recorded as the population of [INSERT WOREDA] according to official administrative data in the
                                                                                                                                                    Census 2007                           99,605                     61,868                       1.19                                                                                                                                                                                           740
Census of 2007 Gregorian Calendaar (1999/2000 Ethiopian Calendar)?

What do you think is recorded as the percentage of rural inhabitants of [INSERT WOREDA] according to official
                                                                                                                                                    Census 2007                            0.84                        0.26                      -2.66                                                                                                                                                                                           726
administrative data in the Census of 2007 Gregorian Calendar (1999/2000 Ethiopian Calendar)?

What do you think is recorded as the unemployment rate of [INSERT WOREDA] according to official administrative
                                                                                                                                                    Census 2007                            4.07                        5.63                       2.21                                                                                                                                                                                           740
data in the Census of 2007 Gregorian Calendar (1999/2000 Ethiopian Calendar)?
Education indicators
What do you think the primary (grades 1-8) enrolment numbers are for [INSERT WOREDA] according to official
                                                                                                                                               EMIS 2006/07-2013/14                       21,929                     14,020                       1.06                       0.61                       0.05                       -12.36                      0.00                       -2.34                      -2.63                       705
administrative data?
What do you think the primary (grades 1-8) pupil-section ratio is for [INSERT WOREDA] according to official
                                                                                                                                               EMIS 2006/07-2013/14                        57.17                      26.71                       5.32                       0.41                       -0.03                                                                                                                                    705
administrative data?
What do you think the primary (grades 1-8) pupils-per-school ratio is for [INSERT WOREDA] according to official
                                                                                                                                               EMIS 2006/07-2013/14                         550                        229                        1.61                       0.57                       0.16                       -5.46                       0.00                       -2.34                      -2.79                       705
administrative data?
What do you think the primary (grades 1-8) pupil-teacher ratio is for [INSERT WOREDA] according to official
                                                                                                                                               EMIS 2006/07-2013/14                        97.61                       354                        15.1                       -0.15                      0.02                                                                                                                                     467
administrative data?
Health indicators
What do you think the Proportion of pregnant women who attended ANC4+ during the current pregnancy is for                                    HSDP Core Plan 2008/09-
                                                                                                                                                                                           0.51                        0.29                       0.08                       -0.13                      -0.53                      -6.81                       0.00                                                                              688
[INSERT WOREDA] according to official administrative data (the baseline of the 2007 Core Plan)?                                                     2013/14
What do you think the Contraceptive Acceptance Rate is for [INSERT WOREDA] according to official administrative                              HSDP Core Plan 2008/09-
                                                                                                                                                                                           0.59                        0.29                      -0.23                       -0.19                      -0.52                      -6.38                       0.00                                                                              688
data (the baseline of the 2007 Core Plan)?                                                                                                          2013/14
What do you think the Rate of Births Attended by Skilled Health Personnel is for [INSERT WOREDA] district                                    HSDP Core Plan 2008/09-
                                                                                                                                                                                           0.33                        0.22                       2.62                       -0.19                      -0.60                      -8.26                       0.00                                                                              688
government according to official administrative data (the baseline of the 2007 Core Plan)?                                                          2013/14
What do you think the Proportion of infants fully immunized is for [INSERT WOREDA] according to official                                     HSDP Core Plan 2008/09-
                                                                                                                                                                                           0.74                        0.22                      -1.24                       -0.08                      -0.57                      -8.06                       0.00                                                                              688
administrative data (the baseline of the 2007 Core Plan)?                                                                                           2013/14

Agriculture indicators
How much land do you think is used for agricultural purposes for [INSERT WOREDA] district government according to                          WCBS Round III (2009/10) &
                                                                                                                                                                                          29,366                     33,892                       6.49                       1.01                       1.01                                                                                                                                     203
data from the WCBS?                                                                                                                           Round V (2012/13)
                                                                                                                                           WCBS Round III (2009/10) &
How much land do you think is used for pastoral purposes for [INSERT WOREDA] according to data from the WCBS?                                                                             27,009                     75,005                       5.97                       1.00                       1.00                                                                                                                                     195
                                                                                                                                              Round V (2012/13)
What do you think is the agricultural income per household per year in [INSERT WOREDA] according to data from the                          WCBS Round III (2009/10) &
                                                                                                                                                                                          11,738                     15,868                       4.58                       0.32                       0.32                                                                                                                                     146
WCBS?                                                                                                                                         Round V (2012/13)
What do you think is the share of households dependent on subsistence agriculture in [INSERT WOREDA] according                             WCBS Round III (2009/10) &
                                                                                                                                                                                           24.29                      23.23                       1.56                       0.92                       0.92                                                                                                                                     132
to data from the WCBS?                                                                                                                        Round V (2012/13)
Revenue indicators
How many tax identification numbers do you think were issued in the last financial year according to data from the                         WCBS Round III (2009/10) &
                                                                                                                                                                                            425                        478                        2.42                       0.33                       0.33                                                                                                                                     146
WCBS for [INSERT WOREDA]?                                                                                                                     Round V (2012/13)
What proportion of income for this district government do you think is received from regional recurrent block grants                       WCBS Round III (2009/10) &
                                                                                                                                                                                           0.81                        0.18                      -1.97                       -0.22                      -0.22                                                                                                                                    190
according to data from the WCBS for [INSERT WOREDA]?                                                                                          Round V (2012/13)
What proportion of income for this district government do you think is received from own sources according to data                         WCBS Round III (2009/10) &
                                                                                                                                                                                           0.20                        0.15                       2.46                       -0.08                      -0.08                                                                                                                                    177
from the WCBS for [INSERT WOREDA]?                                                                                                            Round V (2012/13)

Trade indicators
How many business licenses do you think were issued in the last financial year according to data from the WCBS for                         WCBS Round III (2009/10) &
                                                                                                                                                                                            376                        481                        5.19                       1.36                       1.36                                                                                                                                     177
[INSERT WOREDA]?                                                                                                                              Round V (2012/13)
What do you think is the total revenue collected from issuing and renewing business licenses according to data from                        WCBS Round III (2009/10) &
                                                                                                                                                                                         108,473                     218,841                      8.69                       0.32                       0.32                                                                                                                                     172
the WCBS for [INSERT WOREDA]?                                                                                                                 Round V (2012/13)
Civil Service Rules
What is the amount of regular working hours for a civil servants according to the Civil Service Proclamation?
                                                                                                                                            Civil Servants Proclamation                     39
(Paragraph 32)
What is the allowed amount of annual leave in the first year of service according to the Civil Service Proclamation?
                                                                                                                                            Civil Servants Proclamation                     20
(Paragraph 37, Item 1)
What is the maximum amount of annual leave for a civil servant who has served for more than one year according to
                                                                                                                                            Civil Servants Proclamation                     30
the Civil Service Proclamation? (Paragraph 37, Item 2)
What is the allowed amount of maternity leave according to the Civil Service Proclamation? (Paragraph 41, Item 2)                           Civil Servants Proclamation                     90
What are the types of different disciplinary actions that a civil servant can face for the breach of discipline according to
                                                                                                                                            Civil Servants Proclamation                      6
the Civil Service Proclamation? (Paragraph 67)
What is the notice time required before resigning according to the Civil Service Proclamation? (Paragraph 78, Item 1)                       Civil Servants Proclamation                     30
Notes: The unit of observation is the distict. EMIS is the Education Management Information System of the Ministry of Education, Ethiopia. HSDP is the Health Sector Development Plan of the Ministry of Health, Ethiopia. WCBS is the District government and City Benchmarking Survey, undertaken by the Ministry of Public Service and Human Resource Development and World Bank. Skewness refers to Pearson's moment coefficient of skewness. Column 4 reports the autogregressive coefficient fro
teacher ratio are not displayed due to gaps in the panel series; test statistics and autoregressive parameters are not shows for the demographic indicators as these are from a single observation from the census of 2007; test statistics and autoregressive parameters are not shows for the agriculture, revenue, or trade indicators as these are from a a two-period panel dataset, with insufficent observations to conduct the tests. Figures are rounded to two decimal
places.
Table A3: World Management Survey Indicators
Aggregate Index         Topic                                   Question                                                    Score 1                                                       Score 3                                                                       Score 5

                                                                                                                                                                                                                                      Full set of indicators are tracked formally and continuously. Reviews are
                                     In what kind of ways does your Directorate track how well it is                                                       Directorate tracks a number of performance indicators. These are
   Monitoring        Monitoring                                                                             Directorate does not track performance.                                                                                 conducted regularly and involve representative of all directorate staff groups.
                                     delivering services? Can you give me an example?                                                                               seen and reviewed by senior management only.
                                                                                                                                                                                                                                     The results of the review are formally communicate to all directorate staff.
                                     Are you involved in performance review for your Directorate?
   Monitoring        Monitoring                                                                               Not involved in performance review;                                       Bi-annually                                                                     Monthly
                                     If so, how often does this occur?
                                                                                                                                                           Targets are assigned to the directorate, as well as to the manager
                                     Does your Directorate have a clear set of targets derived from                                                                                                                            Targets are clearly defined for the directorate, manager, and employee levels,
                                                                                                             The directorate does not have defined         and employee levels, and these are generally well understood by
     Other            Targeting      the organization’s goals and objectives? Are they used to                                                                                                                                   and are well understood by all staff. All tasks are directly derived from the
                                                                                                                            targets.                       mid-level staff. However the tasks assigned to staff are not always
                                     determine your work schedule?                                                                                                                                                                 targets, which are regularly reviewed to ensure they remain on track.
                                                                                                                                                                                related to those targets.
                                     When you arrive at work each day, do you and your colleagues
                                                                                                             Staff do not know what their roles and        Staff have a good idea of their roles and responsibilities but it is not Staff have a very good understanding of their roles and responsibilities. Their
     Other            Targeting      know what their individual roles and responsibilities are in
                                                                                                                       responsibilities are.                  always clear how they contribute to their organization’s goals.       own roles and goals are clearly interconnected to those of their organization.
                                     achieving the organization’s goals?
                                     How are targets and performance measures communicated to              Neither targets nor performance measures        Targets and performance measures are formally communicated to           Targets and performance measures are formally communicated and
     Other            Targeting
                                     staff in your directorate?                                                    are communicated to staff.                                 managers and team leaders.                                                  understood by all staff.
                                     How would under-performance be tolerated in your                                                                      Poor performance Is identified through evaluation and is addressed
                    Performance                                                                            Poor performers stay in their positions (no                                                                        Poor performers are identified through regular reviews and are put on a formal
     Other                           Directorate? Can you give me an example of how such a case                                                            through concrete action. Although this applies to most staff, some
                     incentives                                                                                        consequences).                                                                                              performance improvement plan immediately. This applies to all staff.
                                     would be dealt with?                                                                                                               individuals/staff groups get away with it.
                                     Given past experience, have members of [respondent’s                                                                                                                                     Bad behaviour/ breaking the rules is addressed through concrete action. If any
                    Performance                                                                               There are no consequences for bad               Bad behaviour is addressed through concrete action, but the
     Other                           organization] been disciplined for breaking the rules of the civil                                                                                                                          employee breaks the rules, the underlying issues will be identified and
                     incentives                                                                                 behaviour/ breaking the rules.                            underlying issues are not addressed.
                                     service?                                                                                                                                                                                                    rectified. This applies to all employees.
                                     Does your Directorate use performance, targets, or indicators        Staff are rewarded equally (or not rewarded) There is a formal staff evaluation system in place and performance              There is a formal staff evaluation system and performance is rewarded
                    Performance
     Other                           for tracking and rewarding (financially or non-financially) the         irrespective of performance. Individual    is rewarded (financially or non-financially). However, there are no        (financially or non- financially). Rewards are given as a consequence of well-
                     incentives
                                     performance of its employees?                                             performance is not tracked formally                  clear system or criteria for rewarding staff                       defined and monitored individual achievements. This applied to all staff.
                                     When staff in your Directorate are given tasks in their daily        How officers carry out their assignments is       How officers carry out their assignments is jointly decided by the
     Other              Roles        work, how much discretion do they have to carry out their            decided by senior managers. Officers have          officer and senior managers. Senior managers tend to drive the          Officers have complete autonomy in deciding how to carry out their tasks.
                                     assignments? Can you give me an example?                                               no say.                                                      decisions.
                                     Can most staff in your Directorate make substantive                                                                      Staff can contribute to policy formulation and decisions about       Management expects all staff to contribute to policy formulation and decisions
                                                                                                          Staff do not contribute to policy formulation,
     Other              Roles        contributions to the policy formulation and implementation                                                             implementation, but there is no formal forum through which to do       about implementation (formally or informally), and considers this part of their
                                                                                                            nor to decisions about implementation.
                                     process?                                                                                                                  this. Contributions typically only occur when problems arise.                                           duties.
                                                                                                                                                                                                                                   The burden of the directorate’s work is distributed equally among staff. Tasks
                                     Is the workload of achieving your Directorate’s targets evenly                                                          The burden of the directorate’s work is more or less distributed
                                                                                                           A small minority of staff undertake the vast                                                                            are assigned in such a way that the amount of time required and the level of
     Other              Roles        distributed across its different employees, or do some groups                                                            equally among staff. A small minority get away with working
                                                                                                             majority of work within the directorate.                                                                                   difficulty are balanced out so no member of staff finds him/herself
                                     consistently shoulder a greater burden than others?                                                                                     significantly less than others.
                                                                                                                                                                                                                                                                   overburdened.
                                     Thinking about all the projects that your Directorate has been
                                                                                                                                                            Managers try to use the right staff for the right job but do not go to
                                     involved in since your appointment here, would you say that                                                                                                                                   The right staff are always used for a task. Allocation of tasks is based on staffs’
     Other              Roles                                                                                Staff are allocated to tasks randomly.        great lengths to ensure this, or are met with institutional constraints
                                     managers and supervisors try to use the right staff for the right                                                                                                                                                   documented skills and competencies.
                                                                                                                                                                         which may prevent them from doing so.
                                     job?
                                     Does your Directorate make efforts to adjust to the specific                                                                                                                                  The directorate tailors all procedures to the specific needs of its stakeholders.
                                                                                                           The directorate uses the same procedures           The directorate tailors procedures to the specific needs of its
     Other            Flexibility    needs and specific requirements of communities, clients, or                                                                                                                                     The evolution of those needs results in adaptation to plans, project and
                                                                                                                        no matter what.                       stakeholders, but struggles when those needs are complex.
                                     other stakeholders?                                                                                                                                                                                                               policies.
                                                                                                                                                            New ideas or practices are adopted, but in an informal and/ or
                                                                                                                                                                                                                               The adoption of new ideas and practices is an integral part of the directorate’s
                                     How flexible would you say your Directorate is in terms of           New practices are not adopted/ integrated in   isolated manner. The directorate encourages the adoption of new
     Other            Flexibility                                                                                                                                                                                              work. New practices are regularly reviewed and considered, and once adopted
                                     responding to new and improved work practices or reforms?                          the directorate.                   practices, however it is slow to integrate them into its operations
                                                                                                                                                                                                                                             and integrated across the directorate within 6 months.
                                                                                                                                                                                   (more than a year).
                                                                                                             Ad-hoc, no set process for improvement          Existing process to deal with problems Improvements made
                                                                                                             Deal with problems as they arise without                               through meetings                            Exposing problems and suggesting solutions and improvements is part of all
                  Staff involvement/
     Other                           How do problems in your directorate get exposed and fixed?              following an established procedureOnce          Focus on finding solutions, not prevention of future problems          staffs’ daily duty. Continuous improvement is part of the culture of the
                     contribution
                                                                                                          fixed, no further action taken No suggestions      Suggestions from staff involved through meetings (formal or                                          organization.
                                                                                                                             from staff                                                 informal)
                                                                                                                                                                                                                               Staff provide the feedback on which action plans will be based. Focus on both
                  Staff involvement/                                                                                                                    Staff provide feedback in meetings but in an unstructured manner.
     Other                           What kind of feedback do you get in staff meetings?                              No feedback from staff.                                                                                       good and bad performance. Details of the meetings are recorded and
                     contribution                                                                                                                                            Focus on bad performance.
                                                                                                                                                                                                                                                            communicated to all staff.
                                                                                                                                                        Failure can be found in regular meetings (weekly, even monthly for        In addition to 4, tools can be checked up and reported to the manager in
                  Staff involvement/ Let’s say you’ve agreed to a follow up plan at one of your             No action taken. No changes made in the
     Other                                                                                                                                                long-term plans) or at standard points before the deadline. Plans      charge. Meetings (formal/ informal) are held to look into the root causes of
                     contribution    meetings, what would happen if the plan wasn’t enacted?                            operations process.
                                                                                                                                                             can be altered in order to achieve expected results on time.             problems and preventive actions are taken for future similar task.
                                     Do you think the management of your Directorate think about
                                                                                                                                                       Senior management believes that attracting and developing talent Senior management believes that attracting and developing talent is important.
                                     attracting talented people to your Directorate and then doing
     Other             Staffing                                                                           Directorate does not put emphasis on talent is important, but there is no clear system for identifying, attracting or There is a clear system for identifying and attracting talent, developing and
                                     their best to keep them? For example, by ensuring they are
                                                                                                                                                                               retaining such talent.                                                         retaining talent.
                                     happy and engaged with their work.
                                     If two senior level staff joined your Directorate five years ago         No promotion system (no one in the           The promotion system is based on performance. Organization may           Promotion system is based on performance. Organization actively identifies,
     Other             Staffing      and one was much better at their work than the other, would           organization has been promoted for years)          have internal limitations (e.g. few position openings), but do         develops and promotes top performers. Regular assessments, clear set of
                                     he/she be promoted through the service faster?                        The promotion system is based on tenure                everything to get around them (e.g. extra training).               indicators and personalised career plans for individuals (regularly revised).
Table A4: Determinants of Superior Information in Hierarchies
Dependent Variable: Z-score of errors across policy domains
Analysis of Variance (ANOVA) Estimates in column (1); OLS estimates in column (2)
Partial sum of squares and F-statistic p-values in parentheses in column (1); OLS coefficient and clustered standard error at the district-
government-sector level, in parentheses in column (2); F-statistics and associated p-values from the joint test across all categories for
factor variables preceded by "F=" and "P=" respectively in column (2)

                                                                                                                                 (1) ANOVA                               (2) Regression coefficients

Hierarchical structure
District Government [Yes=1]                                                                                                          8.71                                               -0.33
                                                                                                                                    [0.00]                                              [0.14]
Manager                                                                                                                              0.04                                               0.012
                                                                                                                                    [0.83]                                             [0.072]
Sector                                                                                                                               3.61                                               F=0.6
                                                                                                                                    [0.41]                                             P=0.67
Marginal costs
Feature of Claim: Indicator                                                                                                         27.09                                               F=1.92
                                                                                                                                    [0.05]                                              P=0.02
Feature of Claim: Percentile                                                                                                         9.38                                               F=2.73
                                                                                                                                    [0.01]                                              P=0.07
Bureaucrat Characteristic: Years in position                                                                                         0.00                                             -0.00049
                                                                                                                                    [0.95]                                              [0.012]
Bureaucrat Characteristic: Years in organization                                                                                     0.19                                              -0.0022
                                                                                                                                    [0.65]                                             [0.0047]
Bureaucrat Characteristic: Years in civil service                                                                                    0.00                                              0.00013
                                                                                                                                    [0.98]                                             [0.0040]
Bureaucrat Characteristic: Undergraduate education [degree=1]                                                                        0.52                                               -0.069
                                                                                                                                    [0.45]                                               [0.12]
Bureaucrat Characteristic: Postgraduate education [masters=1]                                                                        0.02                                              -0.0092
                                                                                                                                    [0.89]                                              [0.078]
Bureaucrat Characteristic: Gender [female=1]                                                                                         0.53                                                0.046
                                                                                                                                    [0.44]                                              [0.065]
Organization Characteristic: MIS exists                                                                                              0.04                                                0.011
                                                                                                                                    [0.84]                                               [0.18]
Organization Characteristic: Proportion of projects civil servant has adequate
                                                                                                                                     11.74                                              0.01
information on service delivery
                                                                                                                                    [0.00]                                            [0.0052]
Organization Characteristic: Proportion of projects information flows effectively
                                                                                                                                     4.84                                               -0.01
within organizations
                                                                                                                                    [0.02]                                            [0.0057]
Organization Characteristic: Proportion of projects information flows effectively
                                                                                                                                     0.08                                              0.0011
across organizations
                                                                                                                                    [0.77]                                            [0.0055]
Organization Characteristic: Span of control                                                                                         3.85                                               -0.01
                                                                                                                                    [0.04]                                            [0.0053]
Organization Characteristic: Number of principals                                                                                    3.44                                               0.075
                                                                                                                                    [0.05]                                             [0.050]
Marginal benefits
Management practices: Monitoring                                                                                                     4.35                                               -0.15
                                                                                                                                    [0.03]                                             [0.072]
Management practices: Other                                                                                                          3.33                                                0.16
                                                                                                                                    [0.05]                                              [0.11]
Alignment: Mission alignment index                                                                                                   1.10                                                0.14
                                                                                                                                    [0.27]                                              [0.18]
Model                                                                                                                               188.47
                                                                                                                                     [0.00]
Residual                                                                                                                           1239.29
District fixed effects                                                                                                                                         Yes
Adjusted R-squared                                                                                                                     -                                               0.080
Number of observations [clusters]                                                                                                    1455                                            1455 [92]

Notes: Column (1) reports ANOVA partial sum of square estimates. Column (2) reports OLS regression coefficients. Standard errors clustered at the district-government-sector level. Column (2) reports the
standard error in brackets; for factor variables with mulitple categories, the F-statistic of joint test of all category indicators are reported after F= and the p-value is reported below after P=. The unit of analysis
in all columns is the claim. The sample used for the analysis is the set of respondents that make claims about the same districts. Sector fixed effects are indicators for whether the claim is within an agriculture, education, health, reven
table A3. Management practice: Other is the organization average of the average across the z-scores of targeting, incentives, roles, flexibility, staffing, and staff involvement, detailed in table A3. Alignment:
Mission alignment index is the z-score of the row mean of the organization-average response to "To what extent do you believe that your organization's mission is aligned to your own mission?" and the
organization proportion of employees selecting "I am contributing to fulfilling that mission on an everyday basis" to "How consistent with your mission are the various tasks and activities assigned to you on a
day-to-day basis?". Figures are rounded to two decimal places or two significant figures.
                                    Table A5: Decentralization and Information Robustness
                                    Dependent Variable: Z-score of errors across policy domains in columns (1) to (5) and column (8); Rank of errors, in terms of the proportion within distribution, across policy domains
                                    in column (6); Skewness adjusted z-scores of errors across policy domains in column (7)
                                    OLS Estimates

                                    Standard Errors: Clustered at the district government-sector level in column (1) and columns (6) to (8); Clustered at the claimed district government level in column (2); Clustered at the
                                    respondent level in column (3); Clustered at the organization-of-respondent level in column (4); Clustered at the region-sector level in column (5).


                                                                                                                                           (2) Clustering at                                     (4) Clustering at the
                                                                                                             (1) Disaggregating                                        (3) Clustering at                                (5) Clustering at                  (6) Rank of Error             (7) Skewness               (8) Management
                                                                                                                                           claimed district                                         organization-of-
                                                                                                               central dummy                                           respondent level                                region-sector level                    (proportion)              adjusted z-score                controls
                                                                                                                                          government level                                         respondent level

                                    District Government [Yes=1]                                                        -0.31                       -0.33                       -0.33                       -0.33                       -0.33                       -0.10                        -0.39                      -0.28
                                                                                                                      [0.098]                     [0.096]                     [0.064]                     [0.081]                     [0.094]                     [0.027]                      [0.098]                     [0.16]
                                    Federal Government [Yes=1]                                                         0.042
                                                                                                                      [0.091]
                                    Manager                                                                            0.039                       0.034                       0.034                       0.034                       0.034                     -0.00034                       0.029                      0.060
                                                                                                                      [0.067]                     [0.070]                     [0.065]                     [0.068]                     [0.073]                     [0.018]                      [0.059]                     [0.11]
                                    Sector fixed effects                                                               Yes                         Yes                         Yes                         Yes                          Yes                        Yes                         Yes                          Yes
                                    District government fixed effects                                                  Yes                         Yes                         Yes                         Yes                          Yes                        Yes                         Yes                          No
                                    Indicator fixed effects                                                            Yes                         Yes                         Yes                         Yes                          Yes                        Yes                         Yes                          Yes
                                    Percentile fixed effects                                                           Yes                         Yes                         Yes                         Yes                          Yes                        Yes                         Yes                          Yes
                                    Individual controls                                                                Yes                         Yes                         Yes                         Yes                          Yes                        Yes                         Yes                          Yes
                                    Region fixed effects                                                               No                          No                          No                          No                           No                          No                          No                          Yes
                                    Marginal costs and noise controls                                                  No                          No                          No                          No                           No                          No                          No                          Yes
                                    Adjusted R-squared                                                                0.068                       0.069                       0.069                       0.069                        0.069                       0.11                        0.11                        0.057
                                    Observations [clusters]                                                         1455 [92]                   1455 [42]                   1455 [687]                  1455 [122]                   1455 [38]                   1455 [92]                   1455 [92]                   1455 [92]
                                    Federal respondents [percentage]                                                195 [28%]                   195 [28%]                   195 [28%]                   195 [28%]                   195 [28%]                   195 [28%]                   195 [28%]                    195 [28%]
                                    Regional respondents [percentage]                                               234 [34%]                   234 [34%]                   234 [34%]                   234 [34%]                   234 [34%]                   234 [34%]                   234 [34%]                    234 [34%]
                                    District government respondents [percentage]                                    260 [38%]                   260 [38%]                   260 [38%]                   260 [38%]                   260 [38%]                   260 [38%]                   260 [38%]                    260 [38%]
                                    Notes:
rank within the indicator) within the        Clustered
                                      distribution of errors
                                                         standard
                                                             withinerrors
                                                                    an indicator in column (6); Standard
                                                                          are in parentheses.            errors clustered
                                                                                                the skewness-adjusted   z-score
                                                                                                                          at the is
                                                                                                                                  district
                                                                                                                                    the dependent  variable in
                                                                                                                                           government-sector   column (7)
                                                                                                                                                             interacted level
                                                                                                                                                                          - the
                                                                                                                                                                              in skewness
                                                                                                                                                                                 column (1) and columns
                                                                                                                                                                                          adjusted       (6)is
                                                                                                                                                                                                    z-score  tothe
                                                                                                                                                                                                                (8);
                                                                                                                                                                                                                   z-score  of the
                                                                                                                                                                                                                     clustered      absolute
                                                                                                                                                                                                                                at the claimed
                                                                                                                                                                                                                                             error
                                                                                                                                                                                                                                               district
                                                                                                                                                                                                                                                   if the variable
                                                                                                                                                                                                                                                        level      is a proportion
                                                                                                                                                                                                                                                              in column  (2); clustered
                                                                                                                                                                                                                                                                                    variable
                                                                                                                                                                                                                                                                                         at the
                                                                                                                                                                                                                                                                                              (between 0 and
                                                                                                                                                                                                                                                                                                respondent   1), in
                                                                                                                                                                                                                                                                                                           level if the
                                                                                                                                                                                                                                                                                                                     column
                                                                                                                                                                                                                                                                                                                        variable
                                                                                                                                                                                                                                                                                                                             (3); is not a proportion
                                                                                                                                                                                                                                                                                                                                  clustered           and is statistically signific
                                                                                                                                                                                                                                                                                                                                             at the organization-of-respondent
                                    mission on an everyday basis" to "How consistent with your mission are the various tasks and activities assigned to you on a day-to-day basis?"). Noise controls include the time of day of the survey, day-of-survey fixed effects, enumerator fixed effects, an indicator for the
                                    enumerator's subjective assessment of the quality of the interview; an indicator for the duration of the interview (decile of duration fixed effects); controls for the average tenure of the managers in the organization and the proportion of female managers in the organization.
                                    Figures are rounded to two significant figures.
Table A6: Heterogeneity in Effects: Quantile Regression By Absolute Error
Dependent Variable: Z-score of errors across policy domains in columns
Quantile regression estimates
Standard Errors: Clustered at the district government-sector level
                                                       (1) Percentile =       (2) Percentile =        (3) Percentile =       (4) Percentile =       (5) Percentile =        (6) Percentile =       (7) Percentile =        (8) Percentile =       (9) Percentile =
                                                             0.1                    0.2                     0.3                    0.4                    0.5                     0.6                    0.7                     0.8                    0.9
Federal Government [Yes=1]                                  0.0095                   0.045                  0.057                   0.055                  0.071                   0.014                  -0.011                -0.021                   0.066
                                                            [0.034]                 [0.038]                [0.053]                 [0.055]                [0.072]                 [0.078]                [0.092]                 [0.10]                  [0.13]
District Government [Yes=1]                                  -0.08                   -0.15                  -0.21                   -0.21                  -0.20                   -0.21                   -0.27                 -0.40                   -0.26
                                                            [0.039]                 [0.053]                [0.080]                 [0.078]                [0.076]                 [0.088]                 [0.10]                 [0.11]                  [0.14]
Manager [Yes=1]                                              0.019                   0.019                  0.031                   0.038                 0.0074                  -0.014                 0.0035                  0.023                   0.083
                                                            [0.029]                 [0.031]                [0.031]                 [0.032]                [0.051]                 [0.066]                [0.075]                [0.090]                  [0.10]
Sector fixed effects                                          Yes                     Yes                    Yes                     Yes                   Yes                      Yes                    Yes                    Yes                     Yes
District government fixed effects                             Yes                     Yes                    Yes                     Yes                   Yes                      Yes                    Yes                    Yes                     Yes
Indicator fixed effects                                       Yes                     Yes                    Yes                     Yes                   Yes                      Yes                    Yes                    Yes                     Yes
Percentile fixed effects                                      Yes                     Yes                    Yes                     Yes                   Yes                      Yes                    Yes                    Yes                     Yes
Individual controls                                           Yes                     Yes                    Yes                     Yes                   Yes                      Yes                    Yes                    Yes                     Yes
Observations                                                                                                                                             1455 [92]
Notes: Standard errors are in parentheses. Standard errors clustered at the district-government-sector interacted level. Columns report quantile regression estimates for different quantiles of the distribution of the z-score of absolute errors. The unit of
analysis in all columns is the claim. The sample used for the analysis is the set of respondents that make claims about the same districts. The dependent variable is the z-score of the absolute error. Sector fixed effects are indicators for whether the
claim is within an agriculture, education, health, revenue, or trade organization. Percentile fixed effects are indicators for the 75th, 50th or 25th percentile for the respective indicator for the respective region. Indicator fixed effects control for the service
delivery indicator within which the claim is made. Manager is equal to one if the claim is from a manager. Individual characteristics included in the specification: tenure in the civil service of the respondent making the claim; tenure in the organization; an
indicator for whether the respondent is female; an indicator for whether the respondent holds an undergraduate degree; an indicator for whether the respondent holds a masters degree. Figures are rounded to two decimal places or two significant
figures.
Table A7: Decentralisation and Rules

Dependent Variable: Z-score of errors across civil service rules in columns (1) to (9); Rank of errors across civil service rules in column (10); Skewness-adjusted z-score of errors across civil service rules in column (11)

OLS Estimates

Standard Errors: Clustered by organization-of-respondent in columns (1) to (7) and columns (10) to (11); Clustered at the respondent level in column (8); Clustered at the region and sector level in column (9)

                                                                                                                                                                                                                                                                                   (9) Clustering at
                                                                                                                                (3) District                                                                                       (7) Individual         (8) Clustering at                          (10) Rank of Error (11) Skewness
                                                                            (1) Unconditional          (2) Sector FEs                                 (4) Indicator FEs (5) Percentile FEs (6) Management                                                                            region-sector
                                                                                                                              government FEs                                                                                      Characteristics         respondent level                              (proportion)    adjusted z-score
                                                                                                                                                                                                                                                                                         level
Federal Government                                                                  -0.14                   -0.14                     -0.15                    -0.15                  -0.15                     -0.14                    -0.09                    -0.09                    -0.09                   -0.04                    -0.13
                                                                                   [0.042]                 [0.040]                  [0.035]                  [0.035]                 [0.034]                  [0.035]                   [0.035]                  [0.033]                  [0.036]                 [0.010]                  [0.036]
District government                                                               -0.0057                 -0.00058                   -0.011                   -0.011                 -0.012                    -0.011                    0.015                    0.015                    0.015                  -0.014                    0.061
                                                                                   [0.042]                 [0.042]                  [0.039]                  [0.039]                 [0.038]                  [0.038]                   [0.039]                  [0.037]                  [0.039]                [0.0099]                  [0.038]
Management [Yes=1]                                                                                                                                                                                             0.047                   0.00088                  0.00088                  0.00088                   0.011                  -0.0084
                                                                                                                                                                                                              [0.033]                   [0.030]                  [0.032]                  [0.031]                [0.0072]                  [0.028]
Sector fixed effects                                                                No                      Yes                      Yes                      Yes                     Yes                      Yes                      Yes                      Yes                       Yes                     Yes                     Yes
District government fixed effects                                                   No                      No                       Yes                      Yes                     Yes                      Yes                      Yes                      Yes                       Yes                     Yes                     Yes
Indicator fixed effects                                                             No                      No                       No                       Yes                     Yes                      Yes                      Yes                      Yes                       Yes                     Yes                     Yes
Percentile fixed effects                                                            No                      No                       No                       No                      Yes                      Yes                      Yes                      Yes                       Yes                     Yes                     Yes
Individual controls                                                                 No                      No                       No                       No                      No                       No                       Yes                      Yes                       Yes                     Yes                     Yes
Adjusted R-squared                                                                0.0035                  0.0042                    0.012                    0.011                   0.011                    0.011                    0.032                    0.032                     0.032                   0.048                   0.40
Observations [clusters]                                                          6201 [122]              6201 [122]               6201 [122]               6201 [122]              6201 [122]               6201 [122]               6201 [122]               6201 [689]                6201 [38]               6201 [122]              6201 [122]
Federal respondents [percentage]                                                 195 [28%]                195 [28%]               195 [28%]                195 [28%]               195 [28%]                195 [28%]                195 [28%]                 195 [28%]               195 [28%]                195 [28%]                195 [28%]
Regional respondents [percentage]                                                234 [34%]                234 [34%]               234 [34%]                234 [34%]               234 [34%]                234 [34%]                234 [34%]                 234 [34%]               234 [34%]                234 [34%]                234 [34%]
District government respondents [percentage]                                     260 [38%]                260 [38%]               260 [38%]                260 [38%]               260 [38%]                260 [38%]                260 [38%]                 260 [38%]               260 [38%]                260 [38%]                260 [38%]
Notes: Clustered standard errors are in parentheses. Standard errors clustered at the organization level unless stated otherwise. All columns report OLS estimates. The unit of analysis in all columns is the claim. The sample used for the analysis is the set of respondents that make claims about the same districts. The dependent
variable is the mean z-score of absolute errors across all policy domain items claimed by the same respondent about the same district. Sector fixed effects are indicators for whether the claim is within an agriculture, education, health, revenue, or trade organization. Percentile fixed effects are indicators for the 75th, 50th or 25th percentile
for the respective indicator for the respective region. District government fixed effects identify the district that the respondent is making claims about. Indicator fixed effects control for the service delivery indicator within which the claim is made. Management is equal to one if the claim is from a manager. Individual characteristics
included in the specification: tenure in the civil service of the respondent making the claim; tenure in the position; tenure in the organization; an indicator for whether the respondent is female; an indicator for whether the respondent holds an undergraduate degree; an indicator for whether the respondent holds a masters degree. Column
8 clusters the errors at the respondent level. Column 9 clusters the error at the region-sector level where the region is that in which the district being referred to is situated in. Column 10 runs the preferred specification of the model, as per column 7, clustering at the organization levell, but with the dependent variable as the rank of the
absolute error across all policy domains; the rank is represented as the proportion within the distrbution of the errors within the indicator. Column 12 runs the preferred specification of the model, as per column 7, clustering at the organization level, but with the dependent variable as the skewness adjusted z-score of the absolute error
across all policy domains; the skewness adjusted z-score is simply the z-score if the variable is a proportion; if the variable is not a proportion and is statistically significantly skewed relative to the normal distribution (Royston,1991), then the log of the variable is taken and then the z-score is created from the logged values. Figures are
rounded to two significant figures.
Table A8: Decentralization and Management
Dependent Variable: Management practice: Monitoring in columns (1) to (4); Management practice: Other in columns (5) to (8)
OLS Estimates
Robust standard errors
                                                                                                     Management practice: Monitoring                                                                    Management practice: Other

                                                                                                                              (3) Individual                                                                                  (7) Individual
                                                                                   (1)                                       characteristics           (4) Marginal                                                          characteristics          (8) Marginal
                                                                                                       (2) Sector FE                                                        (5) Unconditional         (6) Sector FE
                                                                               Unconditional                                  (organization                costs                                                              (organization               costs
                                                                                                                                average)                                                                                        average)
District Government [Yes=1]                                                           0.13                   0.16                    0.41                    0.54                   -0.12                  -0.10                    0.14                    0.25
                                                                                     [0.15]                 [0.15]                  [0.21]                  [0.21]                  [0.12]                 [0.12]                  [0.14]                  [0.14]
Sector fixed effects                                                                   No                    Yes                     Yes                    Yes                       No                    Yes                     Yes                    Yes
Individual characteristics (organization average)                                      No                    No                      Yes                    Yes                       No                    No                      Yes                    Yes
Marginal costs                                                                         No                    No                      No                     Yes                       No                    No                       No                    Yes
Adjusted R-squared                                                                  -0.0023                 0.034                   0.078                   0.25                   0.00018                 0.049                    0.10                   0.24
Observations                                                                          122                    122                     122                    122                      122                    122                     122                    122
Notes: *** denotes significance at 1%, ** at 5%, and * at 10% level. Standard errors are in parentheses. Robust standard errors. All columns report OLS estimates. The unit of analysis in all columns is the organization. The tiers of government are grouped
into District Government and Centralized Government which incorporates both Federal and Regional tiers, hence the omitted category for tier in the regressions is Centralized Government. The sample used for the analysis is the set of respondents that
make claims about the same districts. The dependent variable is Management practice: Monitoring in columns (1) to (4) and Management practice: Other in columns (5) to (8). Management practice: Monitoring is the organization average of the monitoring
topic, construced from the z-scores of each individual item noted in table A3. Management practice: Other is the organization average of the average across the z-scores of targeting, incentives, roles, flexibility, staffing, and staff involvement, detailed in table
A3. Management practices are in z-scores. Individual characteristics (organization average) include the organization average of: the number of years in current position; the number of years in current organization; the number of years in the service; an
indicator for whether the employee has an undergraduate degree; an indicator for whether the employee has a masters degree; and an indicator for whether the employee is female. Marginal costs include the proportion of managers in the organization that
state that a management information system is in place; the organization average response to "In what proportion of projects would you say information flows effectively between individuals in the same organization?"; the organization average response to "In
what proportion of projects would you say Information flows effectively between organizations?"; the organization average response of managers to "How many personnel do you manage?"; the organization average response of employees to "How many
people would you say regularly give you tasks as part of your formal work duties?"; alignment (the z-score of the row mean of the organization-average response to "To what extent do you believe that your organization's mission is aligned to your own
mission?" and the organization proportion of selecting "I am contributing to fulfilling that mission on an everyday basis" to "How consistent with your mission are the various tasks and activities assigned to you on a day-to-day basis?"). Figures are rounded to
two significant figures.
Table A9: Errors and Appraisal Scores
Dependent Variable: Within-organization z-score of civil servant's total appraisal score
OLS Estimates
Standard Errors: Clustered at the district government-sector level

                                                                                     (2) Tier fixed         (3) Sector fixed        (4) Indicator fixed         (5) Percentile          (6) Region fixed          (7) Individual                                   (9) Management
                                                        (1) Unconditional                                                                                                                                                                (8) Marginal costs
                                                                                        effects                  effects                  effects                fixed effects               effects             characteristics                                       practices

Absolute error (z-score)                                        -0.021                   -0.023                   -0.023                   -0.024                    -0.024                   -0.028                   -0.021                   -0.026                    -0.024
                                                                [0.013]                  [0.013]                  [0.013]                  [0.013]                   [0.013]                  [0.013]                  [0.012]                  [0.012]                   [0.012]
Don't know                                                      -0.085                    -0.12                    -0.13                    -0.14                     -0.14                   -0.096                   -0.083                   -0.031                    -0.040
                                                                [0.066]                  [0.080]                  [0.086]                  [0.088]                   [0.088]                   [0.10]                  [0.092]                  [0.094]                   [0.092]
Tier fixed effects                                              No                        Yes                      Yes                       Yes                      Yes                     Yes                       Yes                      Yes                      Yes
Sector fixed effects                                            No                        No                       Yes                       Yes                      Yes                     Yes                       Yes                      Yes                      Yes
Indicator fixed effects                                         No                        No                       No                        Yes                      Yes                     Yes                       Yes                      Yes                      Yes
Percentile fixed effects                                        No                        No                       No                        No                       Yes                     Yes                       Yes                      Yes                      Yes
Region fixed effects                                            No                        No                       No                        No                       No                      Yes                       Yes                      Yes                      Yes
Individual characteristics                                      No                        No                       No                        No                       No                      No                        Yes                      Yes                      Yes
Marginal costs and noise controls                               No                        No                       No                        No                       No                      No                        No                       Yes                      Yes
Management practies                                             No                        No                       No                        No                       No                      No                        No                       No                       Yes
Adjusted R-squared                                            0.0024                    0.0033                   0.0035                   -0.00049                 -0.00099                  0.018                     0.082                    0.14                     0.14
Observations [clusters]                                      3963 [135]                3963 [135]               3963 [135]               3963 [135]               3963 [135]               3963 [135]                3963 [135]               3963 [135]               3963 [135]


Notes: Clustered standard errors are in parentheses. Standard errors clustered at the district government-sector interacted level. All columns report OLS estimates. The unit of analysis in all columns is the claim of the district government-level official. The sample used
for the analysis is the set of respondents making claims for which there is also evaluaton data for the civil servant. The dependent variable is the within-organization z-score of the civil servant's evaluation outcome as per their annual appraisal. Tier fixed effects control for
whether the respondent works in a federal, regional, or district government office. Sector fixed effects are indicators for whether the claim is within an agriculture, education, health, revenue, or trade organization. Indicator fixed effects control for the specific variable being
claimed over. Individual characteristics includes: years in position; years in the organization; years in the civil service; an indicator for whether the respondent has an undergraduate degree; and an indicator for whether the respondent has a masters degree; an indicator
for whether the respondent is female; and an indicator for whether the respondent is a manager. Marginal costs include the proportion of managers in the organization that state that a management information system is in place; the organization average response to "In
what proportion of projects would you say information flows effectively between individuals in the same organization?"; the organization average response to "In what proportion of projects would you say Information flows effectively between organizations?"; the
organization average response of managers to "How many personnel do you manage?"; the organization average response of employees to "How many people would you say regularly give you tasks as part of your formal work duties?"; alignment (the z-score of the row
mean of the organization-average response to "To what extent do you believe that your organization's mission is aligned to your own mission?" and the organization proportion of selecting "I am contributing to fulfilling that mission on an everyday basis" to "How consistent
with your mission are the various tasks and activities assigned to you on a day-to-day basis?"). Noise controls include the time of day of the survey, day-of-survey fixed effects, enumerator fixed effects, an indicator for the enumerator's subjective assessment of the
quality of the interview; an indicator for the duration of the interview (decile of duration fixed effects); controls for the average tenure of the managers in the organization and the proportion of female managers in the organization. Management practices incude
Management practice: Monitoring and Managament practice: Other. Management practice: Monitoring is the organization average of the monitoring topic, construced from the z-scores of each individual item noted in table A3. Management practice: Other is the
organization average of the average across the z-scores of targeting, incentives, roles, flexibility, staffing, and staff involvement, detailed in table A3. Figures rounded to two significant figures.
Table A10: Errors and Service Delivery
Dependent Variable: Z-Score of service delivery indicator
OLS Estimates
Standard Errors: Clustered at the district government-sector level

                                                                                                                                                                                                                                                                            (9) Skewness-
                                                                                     (1)                                                                                        (5) Individual          (6) Marginal         (7) Management
                                                                                                         (2) Sector FE         (3) Indicator FE         (4) Region FE                                                                                 (8) Rank error       adjusted z-score
                                                                                 Unconditional                                                                                 characteristics              costs                practices
                                                                                                                                                                                                                                                                                of error
Absolute error z-score                                                                 0.011                   0.010                  0.010                   0.011                  0.011                   0.013                  0.016
                                                                                      [0.086]                 [0.086]                [0.086]                 [0.085]                [0.085]                 [0.083]                [0.081]
Don't know                                                                             -0.13                   -0.16                  -0.16                   -0.15                  -0.16                   -0.17                  -0.19                   -0.21                  -0.20
                                                                                       [0.11]                  [0.11]                 [0.11]                  [0.11]                 [0.10]                  [0.10]                 [0.10]                  [0.10]                 [0.10]
Management practices: Monitoring and targeting                                                                                                                                                                                       0.16                    0.16                   0.16
                                                                                                                                                                                                                                    [0.11]                  [0.11]                 [0.11]
Management practices: Other                                                                                                                                                                                                        0.0095                  0.0084                  0.010
                                                                                                                                                                                                                                    [0.11]                  [0.11]                 [0.11]
Rank error (proportion)                                                                                                                                                                                                                                   0.00071
                                                                                                                                                                                                                                                         [0.00075]
Skewness-adjusted z-score of error                                                                                                                                                                                                                                                 0.041
                                                                                                                                                                                                                                                                                  [0.048]
Sector fixed effects                                                                    No                    Yes                      Yes                   Yes                     Yes                    Yes                     Yes                    Yes                     Yes
Indicator fixed effects                                                                 No                    No                       Yes                   Yes                     Yes                    Yes                     Yes                    Yes                     Yes
Region fixed effects                                                                    No                    No                       No                    Yes                     Yes                    Yes                     Yes                    Yes                     Yes
Individual characteristics                                                              No                    No                       No                    No                      Yes                    Yes                     Yes                    Yes                     Yes
Marginal costs and noise controls                                                       No                    No                       No                    No                      No                     Yes                     Yes                    Yes                     Yes
Management practies                                                                     No                    No                       No                    No                      No                     No                      Yes                    Yes                     Yes
Adjusted R-squared                                                                   0.00073                0.0016                  -0.00060                0.041                   0.044                  0.068                   0.075                  0.077                   0.077
Observations [clusters]                                                             2853 [276]             2853 [276]              2853 [276]             2853 [276]              2853 [276]             2853 [276]              2853 [276]             2853 [276]              2853 [276]
Notes: Clustered standard errors are in parentheses. Standard errors clustered at the district government-sector level. All columns report OLS estimates. The unit of analysis in all columns is the claim of the district government-level official. The sample used for the analysis is
the set of respondents at the district government level making claims about their own district governments. The dependent variable is the z-score of the service delivery outcome for the district corresponding to the claim. Sector fixed effects are indicators for whether the claim is
within an agriculture, education, health, revenue, or trade organization. Region fixed effects are indicators for the region of the claim. Indicator fixed effects are indicators for the specific service-delivery indicator in the claim and the outcome. Individual characteristics includes:
years in position; years in the organization; years in the civil service; an indicator for whether the respondent has an undergraduate degree; and an indicator for whether the respondent has a masters degree; an indicator for whether the respondent is female; and an indicator for
whether the respondent is a manager. Marginal costs include the proportion of managers in the organization that state that a management information system is in place; the organization average response to "In what proportion of projects would you say information flows
effectively between individuals in the same organization?"; the organization average response to "In what proportion of projects would you say Information flows effectively between organizations?"; the organization average response of managers to "How many personnel do you
manage?"; the organization average response of employees to "How many people would you say regularly give you tasks as part of your formal work duties?"; alignment (the z-score of the row mean of the organization-average response to "To what extent do you believe that
your organization's mission is aligned to your own mission?" and the organization proportion of selecting "I am contributing to fulfilling that mission on an everyday basis" to "How consistent with your mission are the various tasks and activities assigned to you on a day-to-day
basis?"). Noise controls include the time of day of the survey, day-of-survey fixed effects, enumerator fixed effects, an indicator for the enumerator's subjective assessment of the quality of the interview; an indicator for the duration of the interview (decile of duration fixed effects);
controls for the average tenure of the managers in the organization and the proportion of female managers in the organization. Management practices incude Management practice: Monitoring and Managament practice: Other. Management practice: Monitoring is the
organization average of the monitoring topic, construced from the z-scores of each individual item noted in table A3. Management practice: Other is the organization average of the average across the z-scores of targeting, incentives, roles, flexibility, staffing, and staff involvement,
detailed in table A3. Figures are rounded to two significant figures.
Table A11: Balance Table
Means and standard deviations
                                                                                (1) Control             (2) Treatment                   (3) T-test [1]=[2]
Official's Characteristics
Age                                                                                34.12                      34.04                             0.08
                                                                                   [8.34]                     [8.73]                           [0.52]
Years in position                                                                   2.92                       2.54                             0.38
                                                                                   [3.03]                     [2.34]                           [0.16]
Years in organization                                                               7.69                       7.34                             0.34
                                                                                   [7.10]                     [7.24]                           [0.44]
Years in civil service                                                             12.04                      12.46                            -0.42
                                                                                   [8.25]                     [8.90]                           [0.53]
Number of different organizations in service                                        1.32                       1.46                            -0.14
                                                                                   [1.52]                     [1.97]                           [0.11]
Grade                                                                               7.26                       6.66                             0.60
                                                                                  [12.59]                    [10.22]                           [0.69]
Education [undergraduate degree=1]                                                  0.76                       0.76                             0.00
                                                                                   [0.43]                     [0.43]                           [0.03]
Education [masters degree=1]                                                        0.01                       0.01                             0.00
                                                                                   [0.12]                     [0.11]                           [0.01]
Gender [female=1]                                                                   0.19                       0.17                             0.02
                                                                                   [0.39]                     [0.38]                           [0.02]
Manager                                                                             0.27                       0.26                             0.01
                                                                                   [0.44]                     [0.44]                           [0.03]

Number of bureaucrats                                                               441                        679                             1120
Organizational Characteristics
Number of bureaucrats per organization                                              4.43                       4.44                            -0.01
                                                                                   [0.79]                     [0.91]                           [0.10]
Number of heads per organization                                                    0.95                       0.97                            -0.02
                                                                                   [0.21]                     [0.16]                           [0.02]
Number of managers per organization                                                 0.93                       0.89                             0.04
                                                                                   [0.26]                     [0.32]                           [0.03]
Number of employees per organization                                                2.54                       2.58                            -0.03
                                                                                   [0.73]                     [0.78]                           [0.09]
Ratio of employees to heads                                                         2.56                       2.58                            -0.02
                                                                                   [0.71]                     [0.78]                           [0.09]
Ratio of employees to managers                                                      2.52                       2.61                            -0.09
                                                                                   [0.74]                     [0.71]                           [0.09]
Span of control (employees per manager)                                             5.64                       6.20                            -0.56
                                                                                   [6.06]                     [8.43]                           [0.90]
Number of organizations                                                             127                        196                              323
Notes: Standard deviations are in parentheses. The unit of observation for the first panel is the civil servant and for the second panel the organization. Only
employees and managers responded to the information module and hence only these individuals are used for the statistics in the first panel. These statistics
are based on the survey sample and not on administrative data. The span of control is taken from the survey question "How many personnel do you
manage?"; the total is taken for each organization if there is more than one manager; and the organization-level statistics are reported (each organization has
an equal weight in the `By Organization' statistics). Heads refers to heads of organization or deputy heads of organization. `Manager' refers to managers
within the organization. Grade is the official civil service grade of the professional civil servant, ranging from 1 to 17 in the sample. The 2013-14 National Civil
Service Human Resource Statistics Abstract (Ministry of Civil Service) records the overall female employee percentage to be 35%; this statistic includes
frontline staff. Figures are rounded to two decimal places.
Figure A1: Map of Sampled Districts for Ethiopian Civil Servants Survey




Notes:	The	figure	shows	a	heat	map	of	districts	across	Ethiopia	based	on	the	Census	2007.		The	districts	shaded	in	dark	red	are	those	sampled	in	the	Ethiopian	Civil	Servants	Survey.		The	gray	areas	are	districts	within	the	Somali	region	of	Ethiopia,	which	
was	mostly	excluded	from	the	sampling	frame	of	the	survey	due	to	security	considerations.
Figure A2: Distributions of Management Indices

               .8




               .6
     Density




               .4




               .2



                                                                                                  Management practice: Monitoring
                                                                                                  Management practice: Other
               0
                    -2             -1               0               1                2
                            Z-score of Management Practice

Notes: The figure shows the distributions of the Management practice: Monitoring (black dashed line) and Management practice: Other (gray
solid line). Management practice: Monitoring is the organization average of the monitoring topic, construced from the z-scores of each individual
item noted in table A3. Management practice: Other is the organization average of the average across the z-scores of targeting, incentives,
roles, flexibility, staffing, and staff involvement, detailed in table A3.
Figure A3: Assessing Homogeneity of District Impact

                                                        Homogeneity of District Government Impact in Information
    District government level mean z-score error




                                                   1
              (across all policy domains)




                                                   0




                                                   -1




                                                                                                                   Centralised organisation
                                                                                                                   Decentralised organisations
                                                   -2
                                                         0           10          20           30             40
                                                              Ranking of district government by mean error
                                                                      across all claims (z-scores)

Notes: The figure shows the average claims for centralized (federal and regional) organizations and decentralized (district
government) organizations, averaged by the district government that they are claiming over. The red squares refer to centralized
organizations and the blue circles to decentralized organizations. The sample used for the analysis is the set of respondents that
make claims about the same districts. The error is in terms of z-score within the indicator.