Instruments of Control: Party Leader Endorsements
and Primary Elections in Africa
Donghyun Danny Choi*
Wednesday 15th February, 2017
Abstract
Political parties in the developing world increasingly rely on primary elections to nominate
their candidates for elected office. For party leaders, the introduction of primaries means
losing an important tool with which they can reward and sanction the behavior of elected
party elites. How can party leaders attempt to retain control over the candidate selection
process, even when they no longer possess the institutionalized authority to impose their
choice of candidates for nomination? I argue that party leaders can do so by using a strategy
- political endorsements - to directly appeal to their co-partisans to select their preferred aspirant during the party primaries. Due to the psychological attachment between party leaders and their co-partisans, the party leader’s endorsement acts as a powerful heuristic upon
which primary voters base their evaluation of the aspirant pool. I test this argument using a
series of experiments on likely primary voters of two major political parties in Kenya. I find
that the endorsements and denouncements of a party leader have strong effects on her copartisan’s evaluation of primary aspirants: the party leader’s co-partisans are significantly
more (less) likely to vote for an aspirant endorsed (denounced) by their party leader. The
magnitude of this effect is large enough to offset the effect of well-established predictors
such as candidate quality or performance, and is consistently observed across the partisans
of both parties.
* PhD Candidate, Charles and Louise Travers Department of Political Science, University of California, Berkeley.
[Email: [email protected] ]
I thank Leo Arriola, Ruth Collier, Justine Davis, Thad Dunning, Matthew Gichohi, Jeffrey Paller, Mathias Poertner, Victor Rateng, Lauren Young, the Africa Research Seminar and CPD working group at UC Berkeley, and
the Working Group on African Political Economy (WGAPE) at Stanford University for helpful comments. I am especially grateful to my research team in Kenya: Project managers Stanley Njuguna, Lynne Kazi, and enumerator
supervisors Anastacia Akinyi, Caroline Aziz (Nakuru), Beatrice Marakal, Elphas Ochieng (Kisumu), and 18 enumerators provided excellent assistance in the field. I am also indebted to the NGG Kenya for their support. This
research was approved by the UC Berkeley Office for the Protection of Human Subjects, under protocol number
2016-05-8759. The pre-analysis plan for this project is registered at the Evidence in Governance and Politics (EGAP)
design registry under protocol ID 20160702AA.
1 Introduction
In many political parties across the developing world, party nominations have traditionally
constituted a crucial part of the party leader’s toolkit in exercising control and imposing discipline over the party and its elites (Ostrogorsky, 1902; Field and Siavelis, 2008). The power to
select party candidates has long been concentrated in the hands of the party leader or a select few within the “inner circle”, and such power has been used by party leaders to ensure that
the party’s elected officials remain loyal to them and serve their interests (Ichino and Nathan,
2012). However, as mounting criticism over the lack of internal party democracy places pressure on parties to change, parties are increasingly relying on more open and inclusive methods
of candidate selection such as primaries, where either partisans or members of the voting public choose party candidates by voting in intraparty elections (Öhman, 2004; Carey and PolgaHecimovich, 2006; Ichino and Nathan, 2016).
When parties move to such open modes of candidate selection, how can party leaders attempt to retain control over the party and its elites? In principle, the introduction of primary
elections implies that party leaders must relinquish control over the list of party candidates who
can compete in elections under the party banner. Much of the very recent theoretical and empirical literature on candidate selection in new democracies implicitly or explicitly adopts this
perspective, relegating the role of party leaders and the party leadership to the background:
they focus on factors such as candidate characteristics - candidates’ record of constituency service or activities in the legislature, their performance in campaign debates, and their clientelistic ties with primary voters - and examine how primary voters are likely to respond to these
characteristics (Izama and Raffler, 2016; Ichino and Nathan, 2012, 2013). Yet is the characterization that primary elections lead party leaders to lose their grip over one of the most important
functions of a modern political party - the nomination of candidates - correct? Existing research
has little insight to offer on whether party leaders are able to adapt their strategies in tandem
with the introduction of primaries, and whether these new strategies are in fact effective in influencing the outcomes of primaries.
This paper argues that party leaders are able to strongly influence who gets to represent the
party on the ballot, even when primary elections remove the de jure authority of party leaders
to directly select the pool of party candidates. Specifically, I focus on one strategy frequently
used by party leaders to do so: political endorsements. Building on some very recent agendasetting work that examines how political elites use endorsements to shape electoral processes
in new democracies (Arriola, 2012; Baldwin, 2013; Koter, 2013), I argue that party leaders can
use endorsements and denouncements to inform voters about their preferences over candidates to sway primary voters to choose their allies and do away with their foes (Cohen et al.,
2009).12 I further argue that these endorsements are likely to have an especially strong persuasive effect amongst primary voters in Africa, where partisans have been found to share a
1
While the endorsement of a party leader can feasibly function through other mechanisms to shape nominations - including consolidation of local party cadre support or increased campaign financing directed to endorsed
candidates, the project focuses on the direct effect of endorsements on voters.
2
These endorsements can be for or against a candidate (endorsement versus denouncements), and can take
both overt and subtle forms, where the preferences of the party leader towards the candidate is either made explicit
or is not explicit but can be inferred. A descriptive analysis of what forms these endorsements take, however, is
beyond the scope of this project.
2
strong sense of linked fate and affective attachment with their party leaders (Gichohi, 2016). In
a low-information electoral setting where other commonly-used heuristics such as ethnicity or
party cues have little to no informational value, an endorsement from a trusted source such as
the party leader is likely to act as a powerful heuristic that can influence how voters evaluate
primary aspirants.
In order to test the influence of party leader endorsements on how voters evaluate primary
aspirants and choose among them, I employ and report findings from a series of pre-registered
experimental research designs administered to likely primary voters of the largest incumbent
party (The National Alliance, hereafter TNA) and the largest opposition party (Orange Democratic Movement, hereafter ODM) in Kenya.3
In the first experiment, I examine how party leader endorsements affect the ways in which
primary voters evaluate individual primary aspirants. More specifically, I use a factorial design in which I experimentally varied the the information provided in a radio news segment
on 1) whether the aspirant received a positive or negative (denouncement) endorsement or no
endorsement at all from the real-world party leaders of TNA and ODM, as well as other local
party elites and 2) whether the aspirant’s record on local public service delivery (performance)
is positive or negative. Following the audio treatment, a survey was administered to probe the
respondent’s willingness to support the aspirant in the primaries, as well as a host of other candidate evaluations intended to ascertain potential mechanisms driving the results.4
In the second experiment, I move beyond the single candidate evaluation framework and
probe the effect of party leader endorsements in a setting where voters are faced with a choice
between multiple candidates: I use a discrete choice-based experiment (Hainmueller, Hopkins
and Yamamoto, 2013), in which respondents were asked to choose between the profiles of two
hypothetical aspirants competing in their party’s legislative primaries that differ in terms of
various attributes, including whether or not the candidate received an endorsement from the
party leader.
Evidence from both experiments support the argument that party leaders are able to strongly
influence how partisan primary voters evaluate primary candidates: I find that primary voters
are much more likely to favorably evaluate and vote for a primary aspirant who has received
the backing of the party leader, and significantly less inclined to do so for an aspirant who has
been denounced by the party leader. The magnitude of this effect is substantively large, rivaling or in some cases surpassing the effect of other determinants of primary vote choice such
as candidate performance/quality that existing research has demonstrated to have a substantial impact on electoral outcomes. This effect is consistently observed across the two different
parties studied, and survives adjustments for multiple hypothesis testing. While the design of
the experiments do not allow us to draw strong inferences regarding the mechanisms driving
this effect, I find suggestive evidence from the first experiment that indicates primary voters
consider aspirants endorsed by the party leader to be more loyal to the party and the party
3
The research design and pre-analysis plan for the two experiments are listed in the Evidence in Governance
and Politics (EGAP) registry under protocol ID 20160702AA under the title “Choosing Party Players: Leader Endorsements and Candidate Selection in Africa.”
4
As will be discussed in subsequent sections, the second dimension of the treatment that was experimentally
manipulated - the candidate’s record of local service delivery - was selected because recent experimental research
on voting behavior in Africa has found it to be one of the strongest factors that shape candidate evaluations
(Conroy-Krutz, 2013; Harding, 2015).
3
leader, actively contribute and campaign for the party to take the national seat of power (the
presidency), and be more trustworthy.
The findings presented in this paper have important implications for our understanding of
political parties and democratic accountability in the developing world. First, political parties
in many of these transitional democracies have often fronted primary elections as a crucial
step towards harnessing much-needed “internal democracy” within the party organization and
a building block for institutionalizing the participation of regular party members or voters in
important party decision-making processes (Norris, 1997). However, the findings presented in
this paper suggest that while party elites might have nominally ceded control over candidate
selection processes to the masses to create a semblance of internal democracy, they can still
effectively retain control over who is nominated by the party by adapting their strategies in
accordance with the changes in their rules and procedures. To a certain extent, the level of party
control over nominations is in line with the well-documented influence of political parties over
nominations in more consolidated demacracies (Cobb, 2006; Cohen et al., 2009; Masket, 2009).
The findings also appear to offer additional evidence on the factors that may moderate the
effect of retrospective performance evaluations on voting behavior in new democracies. A large
number of studies have found strong evidence that voters in these countries reward or punish
politicians based on their track record in providing local public goods or clientelistic benefits
(Conroy-Krutz, 2013; Long and Hoffman, 2013; Carlson, 2015; Harding, 2015). While I also find
evidence that suggest voters indeed value politician performance, I also document that these
effects can be strongly moderated, or even counteracted by endorsements. This extends recent experimental work in voting behavior that finds that social identities such as ethnicity can
interact with and moderate the effect of performance-based voting (Adida et al., 2016).
2 Party leaders, Endorsements, and the Democratization of Candidate Selection
2.1 Party Control and the Democratization of Candidate Selection
For many party leaders in new democracies that have long wielded a tremendous amount of
influence over the party apparatus and its elites, the “democratization” of candidate selection
methods poses a fundamental threat to their ability in sustaining dominance over their traditional domain (Cohen et al., 2009, 353). When the authority to select party candidates reside
purely with the party leader and a small number of individuals within the “inner circle”, political
aspirants within the party often have little option but to acquiesce to the will of the leader, for
fear that they will be denied the opportunity to even appear on the ballot (Pennings and Hazan,
2001; Siavelis and Morgenstern, 2012; Ichino and Nathan, 2012). Existing research on candidate
selection has found that control over the candidate selection processes ensures that “aspirants,
candidates, and legislators, will be responsive to the selectorate”, determining where their loyalties will primarily lie (Field and Siavelis, 2008; Hazan and Rahat, 2010). Since it removes the
power to select party candidates from the hands of the leaders, the introduction of mass participation in candidate selection (such as primary elections) can critically undermine the party
leaders’ ability to induce compliance from elected party elites. This in turn opens the door to divided loyalties among intraparty elites as well as lower levels of party cohesion which can have
4
pernicious electoral consequences for the party (Duverger, 1959; Pennings and Hazan, 2001;
Field and Siavelis, 2008; Hazan and Rahat, 2010).
Despite the apparent risks associated with relinquishing control over the candidate selection process, it is often the case that party leaders are compelled to shoulder the risks and open
up the process to mass participation due to other overriding concerns. Existing research, for example, has found that faced with declining popular support or looming electoral defeat, party
leaders often have to adopt party democratizing reforms as a means to re-engage and energize party supporters (Pennings and Hazan, 2001; Hopkin, 2001). Examining the case of Ghana,
recent research has compellingly found that the party leadership bows to pressure from local
party members to introduce primaries in strongholds (nominations in party strongholds are
likely to culminate into a parliamentary seat, and are therefore likely to be more competitive),
fearing that denying local party elites the opportunity to extract rents from primary aspirants
may have negative implications for their party’s electoral performance (Ichino and Nathan,
2012).
What, then, does the (voluntary or involuntary) adoption of mass participatory candidate
selection mean for the party leader’s ability to retain control of the candidate selection process, and thereby, the party elites in elected office? Somewhat understandably, the dominant
narrative within the existing literature on the consequences of primary reforms has portrayed
party leaders as passive actors who do little to nothing to fight the consequences of mass-driven
candidate selection. Many of these studies begin from the premise that once primaries are
introduced, the choice over party candidates is entirely determined by the preferences of the
electorate participating in the primaries, and proceed to examine the effect of candidate characteristics such as candidate performance in constituency service or campaigning on primary
voters (Izama and Raffler, 2016; Ichino and Nathan, 2016).
While these studies have combined innovative methods and novel data to examine primaries in the context of the developing world, overemphasis on the bottom-up factors shaping
candidate selection seems to have underplayed the countervailing influences that party leaders can have in retaining their hold of party nominations even when they no longer have the de
jure authority to exercising control (Cohen et al., 2009; Masket, 2009). Party leaders are first and
foremost politicians themselves, strongly motivated by their desire to remain in power. Given
that the tenure of party leaders at the helm of the party is likely to be contingent on her ability to
“whip” party elites into remaining loyal to them, party leaders have a strong incentive to insert
themselves into candidate selection processes so that they can make sure loyalists are nominated and enemies are deselected. Evidence from earlier reforms adopted by Western European parties suggest that party leaders are often able to retain ultimate control over candidate
selection despite the adoption of primary elections, and install candidates that they support
(Hopkin, 2001; Katz, 2001).
2.2 Leader Endorsements, Partisan Primary Voters, and Primary Elections
How can party leaders attempt to retain control of the candidate selection process, even with
the introduction of mass-participatory selection methods such as primaries? Amongst the many
ways in which a party leader can attempt to influence the outcome of candidate selection, I consider one important strategy that has been frequently used by party leaders in new democracies:
political endorsements. Building on the recent agenda-setting work on the electoral implica5
tion of elite endorsements in new democracies (Arriola, 2012; Baldwin, 2013; Koter, 2013), I
argue that party leaders can often directly appeal to primary voters by using endorsements to
signal his preference over aspirants competing in party primaries.
Both the extensive use of this direct appeal strategy and its potential effectiveness in influencing primary voters are related first and foremost to the low-information conditions often
associated with electoral competition in new democracies (Chandra, 2007; Ferree, 2010). In environments where there is little differentiation in terms of ideological or policy positions across
both parties and candidates, voters are often compelled to rely on heuristic shortcuts that allow them approximate the behavior of the fully-informed voter (Lupia, 1994; Lau and Redlawsk,
2001). Yet voting “correctly” is made more challenging in the context of primary elections because heuristic shortcuts that voters can rely on often become uninformative or irrelevant: first,
party cues that traditionally act as powerful sources of information in elections with interparty
competition is by definition rendered uninformative in the context of intraparty competition
(Arceneaux, 2008; Sniderman and Stiglitz, 2012; Boudreau and MacKenzie, 2014). Furthermore,
the usefulness of co-ethnicity , which has been found to be an important informational shortcut
in multi-ethnic societies (Ferree, 2010; Arriola, Choi and Gichohi, 2016), is diminished in lowertier elections such as legislative primaries because voters are seldom faced with a multi-ethnic
aspirant pool with candidates who are members of ethnic outgroups (Carlson, 2015).
Under circumstances in which widely-used heuristics are of limited utility, primary voters
are obliged to seek alternative cues that they can rely on to choose the party candidate. The
party leader’s endorsement (or denouncement) regarding primary aspirants becomes a highly
persuasive alternative for primary voters in these contexts because of the privileged relationship forged between party leaders and their partisans in many patronage-based democracies
(Chandra, 2007; Van de Walle, 2003). Political parties in these countries are seldom organized
based on programmatic or ideological differences - rather, parties are often formed based on
existing social cleavages across ascriptive identities such as race, ethnicity, or religion (Madrid,
2012; Elischer, 2013). When ascriptive loyalties become the basis of electoral mobilization in
patronage democracies, elections, especially for national office such as the presidency, become
a contest between these groups to secure future access to state resources. In this regard, party
leaders often become synonymous with political representative of the identity groups, whose
electoral success is likely the most important factor that structures the extent to which these
groups will or will not benefit from the spoils of holding the national seat of power (Van de
Walle, 2003, 2007) .
Due to the link between the political success of party leaders and the material fate of the
party (and thereby the identity groups that the party represents), partisans are likely to develop strong psychological attachments to the party leader herself, separate from the party.
As Gichohi (2016) demonstrates in the context of Kenya, individuals demonstrate a high sense
of linked fate not just towards members of their groups, but to the political leaders of their
groups. This psychological attachment or sentiment towards the political leader is likely to
endow heuristic value to the opinion of the leader, upon which voters may rely in making political decisions, including when they participate in party primaries to select the party candidate
(Dawson, 1994). The party leader’s preferences over the pool of aspirants will factor in significantly when voters select the candidate precisely because they perceive the party leader’s preference and interest to be intertwined with their own. Rather than engage in the effort-intensive
task of acquiring information pertinent for the evaluation primary aspirants, partisans are likely
6
to take the endorsement or denouncement of party leaders as an accurate assessment of how
well-aligned or misaligned aspirants are with their interests.
These arguments yield a number of testable predictions for the expected effects of party
leader preferences over primary election outcomes. First, when primary voters are provided
information on whether the aspirant was endorsed or denounced by the party leader, they are
likely to adjust their evaluation of the aspirant so that it aligns with the opinion of the party
leader. Second, because the strength of the endorsement heuristic, primary voters are likely to
discount the value of other types of information that might otherwise affect their assessment of
the aspirants, such as candidate quality or performance. These predictions constitute the main
hypotheses to be tested in the empirical components of the paper.
3 The Case: Political Parties and Candidate Selection in Kenya
3.1 Political Parties and Electoral Politics
Kenya has held five simultaneous presidential and parliamentary elections since the reintroduction of multiparty elections in 1992, when the ruling Kenya African National Union (KANU)
lifted the ban on opposition political parties from competing in elections. While elections in
Kenya have gained the reputation of being highly controversial affairs, the credibility of which
has been frequently questioned by international and domestic observers, most agree that the
parliamentary elections have been conducted with much higher levels of integrity and order
(Cheeseman, 2008). Under the new constitution that was implemented in 2013, there are a
total of 349 members of the lower house of parliament (otherwise known as the national assembly), 290 of which are elected from single member district constituencies, with another 47
women representatives elected from each of the counties. The upper house (otherwise known
as the senate) is comprised of 67 members, 47 of which are elected directly by the voters in the
county under SMD rules.
Since the democratic transition, Kenya’s party system has been characterized by high levels of fragmentation and volatility (Khadiagala, 2010). Political parties that emerged from the
aftermath of KANU dominance were organized along ethnic interests, often revolving around
powerful personalities whose status as ethno-regional kingpins enabled them to command an
immediate en-bloc following amongst their coethnics (Elischer, 2013). The inability to win an
outright victory in the presidential elections solely based on the support of their ethnic group
has often forced parties to pursue pre-electoral coalitions that temporarily bring together party
leaders and their ethnic group’s interests based on a post-election power-sharing scenario (Arriola, 2012). This duality - ethnic parties cum multi-ethnic coalitions - defines the Kenyan political landscape to this day. In the most recent 2013 general elections, the Jubilee Coalition,
comprised of Uhuru Kenyatta’s The National Alliance (TNA) and William Ruto’s United Republican Party (URP) among others, took the presidential election in the first round against the
Coalition for Reforms and Democracy (CORD), comprised of Raila Odinga’s Orange Democratic
Movement (ODM) and Kalonzo Musyoka’s Wiper Democratic Movement (WDM). The Jubilee
coalition also managed to take the majority in both chambers of parliament, taking 167 seats
in the National Assembly and 30 seats in the Senate. Although the CORD coalition formed the
minority in parliament (141 seats in the National Assembly, and 28 seats in the Senate), as an
7
individual party ODM managed to win the largest number of seats for the National Assembly
with 99 seats, and tied for first in the senate with 17 seats.
3.2 Candidate Selection for Legislative Office
Party nominations for legislative office seem to have been a source of much controversy since
the very early days of post-independence Kenya. During a time in which the only real electoral
competition that KANU faced was the challenge posed by Jaramogi Oginga Odinga’s KPU, much
anxiety amongst sitting members of parliament of KANU, and especially the backbenchers, over
the possibility that the party leadership could unilaterally “deselect them out” of the party fomented elite pressure to introduce primary elections. (Hyden and Leys, 1972; Widner, 1993).
Following the Ol Kalou declaration of 1969, in which 19 KANU MPs formally requested primary
elections be adopted as KANU’s mode of candidate selection, President Jomo Kenyatta committed himself and the party to organizing primary elections for the selection of KANU candidates.
The subsequent KANU primaries throughout the era of single party rule seem to be relatively
well-regarded, competitive affairs in which anyone who could pass the nominal screening of
the party apparatus could compete (Hyden and Leys, 1972; CGD, 2005).
The same characterization cannot be made of party nominations of various political parties during the early periods of the multiparty era. Academic and journalistic accounts of party
nominations since 1992 unequivocally demonstrate that processes laid out in party constitutions or governing documents for party nominations were seldom adhered to in practice (CGD,
2005; Cheeseman, 2008). Party leaders were central in these subversions of procedure: nominations were often conducted at the whim of the party leaders and a small number of cronies
in the inner circle of the party leadership, with nomination certificates frequently being issued
to aspirants who could demonstrate their loyalty to the party leaders (Khadiagala, 2010; Willis
and Chome, 2014). Even in settings where parties “held” primaries, the process was plagued
by lack of organization and management, allegations of an unlevel playing field and favoritism,
and even disruption in the voting process due to violence (CGD, 2005).
The state of affairs improved somewhat for the nominations conducted prior to the 2007
and 2013 general elections. Starting with the 2007 elections, parties on both side of the aisle
started holding primaries in their strongholds and swing constituencies where there were multiple aspirants competing for the party ticket. While not without problems, these were considered some of the earliest attempts to legitimately adopt primary elections to select party
candidates (Cheeseman, 2008). The trend continued in the 2013 elections, when TNA, URP
(Jubilee coalition), ODM, and to somewhat lesser degree WDM (CORD coalition) implemented
primaries as the norm rather than the exception in most of the constituencies for which they
fielded candidates.
However, the implementation of primary elections does not mean that the party leaders
relinquished their influence over who would eventually be selected as the party’s candidate.
While there were highly-publicized defeats suffered by the party leaders’ allies during the party
primaries in both 2007 and 2013 (Lynch, 2014), most of their closest allies managed to secure the
party ticket. Numerous media accounts on party nominations during the 2013 elections suggest that party leaders were able to tap into their privileged relationship with their co-partisan
supporters and use a carefully-calculated combination of explicit and implicit endorsements
8
for aspirants closest to them.5 These journalistic observations are largely consistent with the
accounts of aspirants who competed in the parliamentary primaries for the major political parties in 2007 and 2013. A parliamentary aspirant who unsuccessfully contested in the nomination for one of the major opposition parties, in recounting what happened in the run-up to the
primaries said:
“He (the party leader) all but endorsed one of the other candidates who was close
to him and was holding an important position in the party. After that, it was almost
impossible to campaign because everybody knew that this man had the support of
the leader.6 "
Even eventual winners in the primaries were willing to concede that the backing of the party
leader was one of the most important factors that contributed to their victory in primaries:
“... so when voters saw my development credentials and also saw the my party
leader was with me, they wanted me to get the party ticket.7 "
“ ... I would be lying if I said it did not matter. ... but MPs must work closely with the
party leader ... It is important that I have a close relationship with him.8 "
Furthermore, many politicians agreed that the party leader’s endorsement is an important factor that will contribute to an aspirant’s success in the party primaries even in future elections.
Out of the 34 former aspirants and sitting members of parliament that were asked, 29 of them
(85%) answered that they expected that endorsements would play an important role for the
success of a primary aspirant.
4 Study Design and Sample Selection
4.1 Study Design and Respondent Recruitment
To examine the effect of party leader endorsements on partisan primary voters, I embedded two
separate experimental studies within the context of a large scale voter survey conducted in four
parliamentary constituencies in Nakuru and Kisumu county, Kenya.9 The survey was administered on a total of around 2400 self-identified likely TNA (incumbent) and ODM (opposition)
primary voters in Nakuru and Kisumu, using hand-held electronic tablet devices. The two experimental designs embedded in the survey were chosen based on their respective strengths
and weaknesses, and how the results from each would complement each other in drawing
strong inferences regarding the effect of party leader endorsements. The details of each experimental design are presented in subsequent sections.
5
See for example, “Oduol is a project and an Enemy, Says PM”, The Star, February 8, 2013; “URP Aspirants
in ‘Panic Campaigns’ Ahead of Nominations”, The Star, January 10, 2013; “Othaya MP Aspirants Oppose Kibaki,
Uhuru Endorsement”, The Star, December 28, 2012
6
Transcript of interview with an opposition aspirant, Interview Subject 2015-PO32, conducted on May 15, 2015.
7
Transcript of interview with an Incumbent aspirant, Interview Subject 2015-PI4, conducted on Feb 8, 2015.
8
Transcript of interview with an opposition aspirant, Interview Subject 2015-PO21, conducted on March 24,
2015.
9
Nakuru town’s two parliamentary constituencies were Nakuru Town West and Nakuru Town East. Kisumu
town’s two parliamentary constituencies were Kisumu Central and Kisumu East. The respondents were drawn
from all four of these constituencies to maximize the geographic coverage within these locales.
9
Figure 1: Study Sites - Nakuru and Kisumu County
(a) Nakuru Town West and East Constituency
(b) Kisumu Central and East Constituency
Note: Highlighted areas denote the two constituencies sampled within the respective counties. In Nakuru county
these were Nakuru Town East and Nakuru Town West constituencies. In Kisumu, these were Kisumu Central and
Kisumu East constituencies
Respondents were recruited through door-to-door canvassing using a random-walk method
modeled after the Afrobarometer protocol for household survey sampling. Within each constituency, estates to be surveyed were chosen at random after being listed in pairs. Enumerators started from previously selected landmarks within the chosen estate and executed the walk
method to identify households where interviewing would begin.10 In each household, enumerators followed the Kish grid method to determine which individual, over the age of 18, would
be interviewed. After the respondent was identified, they were administered a short screening
questionnaire that determined eligibility. Only those who were residents of the constituency,
was either i) a registered party member or ii) reported that they had a close attachment towards
either of the two parties, and iii) reported that they were likely to participate in the upcoming
2017 party primaries were eligible to participate in the survey, pending their consent. Respondents received a mobile phone airtime voucher worth 100 Kenyan shillings (roughly equivalent
to $1.00) after completing the interview as compensation for their time. After each successful interview, enumerators skipped a predetermined number of households and repeated this
process until the day’s target was reached.
10
The exact walk protocol used varied by day. The details of the protocol are available upon request to the author.
10
Table 1: Sample Descriptive Statistics (N=2392)
Variable
Female
Voted in last election
Registered party member
Number of years in constituency
Current living conditions
(1=very bad)
Presidential approval
(1=completely disapprove)
Opposition leaders approval
(1=completely disapprove)
Party feeling thermometer
Party leader feeling thermometer
Party leader job approval
Party leader linked fate
Min
Max
Mean (SD)
0
0
0
0
1
1
1
68
0.51 (0.50)
0.81 (0.39)
0.62 (0.49)
14.66 (12.20)
1
5
3.31 (1.03)
1
7
4.71 (1.99)
1
7
4.43 (2.07)
0
0
1
1
100
100
7
4
76.20 (17.92)
79.00 (18.66)
5.90 (1.17)
2.60 (1.21)
4.2 Sample Characteristics
The resulting sample yields a total of 2392 likely incumbent (TNA) and opposition (ODM) primary voters. Table 1 presents the descriptive statistics on the final sample. The sample is almost
exactly balanced on gender. By virtue of sampling self-identified likely primary voters, respondents seem to be active participants in politics, and have favorable evaluations of their political
parties and their party leaders: 81% of the sample report having voted in the previous general
election. 62% self-identified as registered party members. The average rating on the feeling
thermometer for the political party is 76/100, while for the party leader, it is slightly higher at
79/100. Respondents on average strongly approve (6 on a 7 point likert scale) of their party
leader’s job performance.
Additionally, the sample is almost evenly split in terms of the number of incumbent and
opposition supporters: 1205 out of the 2392 reported identifying with the ODM (opposition)
and 1188 were aligned with TNA (incumbent). The two largest ethnic groups represented in the
sample are the Luo (39.1%) and Kikuyu (35.9%), who constitute the main ethnic support base
for the ODM and TNA respectively. These two groups are followed by the Luhya (11.1%), Kisii
(4.3%), Kalenjin (2.6%), and Kamba (2.0%).
5 Experiment 1: Candidate Evaluation Experiment
5.1 Experimental Design and Methods
In testing the hypothesized effect of party leader endorsements and denouncements, I begin by
examining whether they influence voter evaluations of a single primary aspirant or candidate:
any attempt to connect party leader endorsements to potential changes in vote choice remains
tenuous without establishing whether and how they change voter perceptions of candidates. I
therefore conducted a simulated radio news experiment in which I experimentally manipulated
11
Table 2: Treatment Assignment Matrix for Radio News Experiment
Candidate Performance
Endorsement
Leader Endorses
MCAs Endorse
Leader Denounces
MCAs Denounce
No opinion
High
Low
(1) N=227
(2) N=220
(3) N=257
(4) N=262
(5) N=239
(6) N=219
(7) N=247
(8) N=255
(9) N=215
(10) N=252
the information provided to likely TNA and ODM primary voters regarding a fictitious aspirant
seeking the party nomination for the local parliamentary seat.
The content of the audio news segment that delivered this information varied in terms
of whether i) the aspirant received an endorsement from their respective party leaders (Raila
Odinga for ODM and Uhuru Kenyatta for TNA), other elected local party elites (members of
county assembly) or did not receive an endorsement at all and ii) the aspirant’s performance on
providing services to the constituency.11 The addition of the second dimension manipulated
in the treatment was in response to the emerging narrative in the recent literature in African
voting behavior: that voters in African democracies are “performance voters” that privilege the
candidate’s credentials on her ability to deliver local public goods and particularistic benefits
(Conroy-Krutz, 2013; Weghorst and Lindberg, 2013; Harding, 2015).12 13
The experiment follows a 5 x 2 factorial design. For the leader endorsement/denouncement
component of treatment, five different levels were assigned: first, where the party leader purportedly endorses the candidate, second, where party-affiliated local politicians (MCAs) endorse the candidate, third, where the party leader purportedly denounces the candidate, fourth,
where party-affiliated local politicians (MCAs) denounce the candidate, and finally fifth, where
there is no information on whether the candidate was endorsed.14 The no endorsement condi11
In order to mitigate concerns of order effects, I randomized the order in which the two dimensions were presented in the audio file. The results presented hereafter are not vulnerable to the order in which the information
was provided.
12
While the dominant narrative in the African voting behavior literature is ethnic voting (see Adida (2015) for
a recent experimental analysis on the effect of coethnicity), I do not compare the effect of leader endorsements
against candidate coethnicity with the respondent because of the geographically clustered nature of ethnic groups
in Africa. Carlson (2015) argues that voters in Africa are highly “unlikely to encounter non-coethnic candidates
in races for subnational office.” Given that the experiment is running in the context of a parliamentary primary
contest in Nakuru and Kisumu, two towns inhabited by a majority of ethnic Kikuyus and Luos respectively, the
aspirant portrayed in the audio clips are set to be Kikuyu and Luo respectively - i.e. the majority ethnic group of
that region.
13
By comparing the effect of endorsements against the candidate’s performance on constituency service, I subject it to a hard test against a candidate attribute that has been repeatedly demonstrated to have a statistically
significant and substantively large effect on voters. It is also a particularly hard test since existing literature on the
effect of elite endorsements centers around the idea that endorsements affects voters by altering their expectations
about the politician’s performance on service delivery (Baldwin, 2013)
14
I include this MCA level to the endorsement treatment to differentiate between the effect of party leader endorsement and an endorsement by other lower-level party politicians. Comparing the effect of party leader endorsements or denouncements to those issued by other political actors prevents us from conflating the effect of a
12
tion was included as the reference category. For the performance arm of the treatment, two levels were assigned: one where the candidate has a positive development record (high candidate
performance), or a negative development record (low candidate performance). Respondents
were randomly assigned to one of the 10 treatment conditions until the target total sample size
of 2,400 was reached across the two locations (1200 respondents in Nakuru, 1200 respondents
in Kisumu). A tabular presentation of the treatment categories and the actual number of respondents assigned to each category is included in Table 2. Respondents were exposed to the
experimental treatment on an electronic tablet device either in the language of their choice:
Swahili or English. The subjects were then asked a battery of post-treatment questions measured on a 7 point scale, including our main outcome of interest “How likely are you to vote for
the candidate in the TNA/ODM party primaries?”.
The experimental design adopted for this study falls short of the full realism that is considered to be the defining characteristic of field experiments: the subjects were given information
about a fictitious aspirant (albeit with an endorsement from powerful real world party leaders),
and their responses were self-reported vote intentions.15 However, a number of precautions
were taken to enhance the reality of the experimental intervention. Specifically, to mirror the
way in which information about aspirant candidature is conveyed and presented to voters in everyday life, the news segment was modeled after typical news coverage of political candidates
and campaigns by national and local radio news stations: this included a locally-hired actor
narrating the news script, adopting the tone and accent of a local news anchor in-so-doing, as
well as professional editing that added audio-acoustic effects to enhance the reality of the news
segment.16 In addition, respondents were only debriefed at the end of the survey as to the fictitious nature of the aspirant portrayed in the news segment. The precautions taken seem to
have had the desired effect: enumerators report that respondents perceived the aspirant to be
a real contender in the party primaries.17
party leader endorsement with the effect of any endorsement regardless of the identity of the endorser.
15
The decision to use a fictitious aspirant was made in response to two concerns: first, given the perceived influence of party leaders, providing partisans with information about the nature of the relationship between party
leaders and an existing aspirant raised concerns about provoking reactions from other aspirants, and ultimately
influencing the party primaries. Second, given that the relationship between incumbent politicians and the party
leader are quite often discussed in the local media, partisans often hold priors about this relationship. Short of
using outright deception, it would have been challenging to claim that the party leader has both endorsed and
denounced the same aspirant. The use of fictitious aspirant (coupled with real party leaders) addresses both concerns over interfering in elections, and enabling the clean manipulation of endorsement or denouncement status
of the aspirant without the use of deception.
16
Audio-visual treatments have recently been used with some success for experimental research in Africa (McCauley, 2014; McClendon and Riedl, 2015), including in my prior work with coauthors (Arriola, Choi and Gichohi,
2016). Post-survey reports from enumerators indicate that significant proportion of respondents perceived the
candidates portrayed in the news segment to be real contestants in the upcoming party primaries. I opted for
an audio rather than a video treatment to minimize the possibility that experimental results would be subject to
heterogeneity induced by the perceived difference in the delivery of treatments.
17
Prior to being debriefed, respondents frequently exhibited behavior during the interview that suggested that
they were rendering judgements about an aspirant they perceived to be real. More than 10% of the respondents
asked follow up questions about the aspirant, including additional background and contact information of the
aspirant. In more than 50 occasions, respondents extended an invitation to the aspirant to attend the local ward /
village meetings to address constituents.
13
5.2 Results: Intention-to-Treat Analysis
Do the endorsement and denouncement of party leaders affect primary voter’s evaluation of
aspirants? In accordance with the pre-registered analysis plan, I take an intention-to-treat analysis approach, where I simply compare the average responses among respondents assigned to
each treatment and control condition. While this approach identifies the causal effect of treatment assignment, experimental non-compliance makes it highly likely that the results from this
analysis is an underestimate of the treatment effect. I therefore also estimate the complier average causal effect (CACE) for the main results.
Table 3: Party Leader Endorsements / Denouncements and Primary Vote
Intentions
Leader Endorsement Effects
Leader Denouncement Effects
(1)
(2)
(3)
(4)
(5)
(6)
ITTa
SEb
Control
0.953
(0.105)
4.334
1.115
(0.127)
4.312
0.817
(0.163)
4.356
-0.647
(0.114)
4.334
-0.808
(0.144)
4.312
-0.486
(0.177)
4.356
Sample
N
R2
Pooled
937
0.080
TNA
449
0.138
ODM
488
0.049
Pooled
1003
0.031
TNA
506
0.058
ODM
497
0.015
Leader Endorsement Effects
Leader Denouncement Effects
(7)
(8)
(9)
(10)
(11)
(12)
CACEc
SEd
Control
1.110
(0.124)
4.334
1.269
(0.145)
4.312
0.970
(0.197)
4.356
-0.858
(0.150)
4.334
-1.029
(0.182)
4.312
-0.672
(0.243)
4.356
Sample
N
R2
Pooled
937
0.054
TNA
449
0.130
ODM
488
0.018
Pooled
1003
0.052
TNA
506
0.072
ODM
497
0.036
a
b
c
d
Estimated Average Intention-to-Treat Effects (ITTs) of the party leader endorsement
/ denouncement treatments on the main outcome (vote intention in primary elections), pooling across performance dimensions. ATEs are estimated against pure controls in which no endorsement information was provided.
Robust standard errors (SEs) from linear regression analysis.
Complier Average Causal Effects (CACEs) are estimated using Two-Stage least squares
(2SLS) regression in which the first stage regresses the compliance status indicator
variable against the treatment assignment indicator.
Robust standard errors (SEs) from 2SLS regression analysis.
Leader Endorsement/Denouncement Effects
Table 3 presents the main findings for the second experiment. In columns (1)-(3), I present
the estimated intention-to-treat effects (ITTs) of the party leader endorsement treatments on
the main outcome (primary vote intention) vis-à-vis the pure control conditions for the pooled
14
sample as well as the sample disaggregated by party (TNA and ODM). Columns (4)-(6) similarly
presents the estimated ITTs of the party leader denouncement treatments. The first row of each
panel in Table 3 presents the ITTs while the second presents estimated robust standard errors
from linear regression.
I find robust evidence that party leader endorsements and denouncements have a strong
effect on partisan evaluation of primary aspirants. As seen in Column (1), the average primary
vote intention for the aspirant who is endorsed by the party leader is almost a full point (0.953)
larger than the average vote intention of an aspirant who has neither been endorsed nor denounced by the party leader. The primary vote intention of an aspirant who, on the other hand,
has been denounced by the party leader is 0.64 points smaller than that of an aspirant who has
neither been endorsed nor denounced. Both of these differences are statistically significant
at p<0.001, and survive the Benjamini-Hochberg correction for multiple testing at an FDR of
0.05.18
One of the inferential concerns in the experiment is that subjects may not perceive themselves to be in the intended treatment condition to which they were randomly assigned. For
example, even when a respondent was exposed to a radio news segment in which the primary
aspirant was endorsed by the party leader, they might perceive the aspirant to be denounced
by the party leader. Imperfect compliance is a cause for concern because I am interested in
estimating the effect of respondents perceiving aspirants to be supported or denounced by the
party leader on the primary vote intentions. In the post-treatment survey, I embedded manipulation check questions designed to probe whether respondents could correctly identify (or
recall) the aspirant portrayed in the news segment was endorsed, denounced, or received neither endorsement nor denouncement from either the party leader or local MCAs. While I do
not present the detailed results for the manipulation checks, on average around 75-85% of respondents were able to correctly recall the information regarding the aspirant portrayed in the
radio news segment.
Given that non-compliance tends to dilute the effects of treatment, I estimate the complier
average causal effects (CACEs) using the standard instrumental variables approach in which I
use the assignment to treatment status as an instrument for actual treatment receipt. Results
from the Two-Stage least squares (2SLS) regression in which the first stage regresses the treatment receipt indicator against the assignment to treatment indicator are reported in columns
(7)-(12) in Table 3. As expected, the size of the local average treatment effect of party leader
endorsements amongst compliers are appreciably larger than the estimates of the intentionto-treat analysis: the CACE for the leader endorsement effect is 1.110 on a 7 point scale, around
16% larger than the ITT estimate. The difference between the ITT and CACE estimates are larger
for the denouncement effect: the CACE is 0.211 points larger, or 33% larger than the ITT.
The magnitude of the party leader endorsement and denouncement effects are substantively large and important. To illustrate the substantive changes, Figure 2 plots the average
evaluations for the primary aspirant based on whether the party leader endorsed the aspirant,
denounced the aspirant, or there was no endorsement offered, pooling across the performance
18
Since the main outcome is measured on a 7 point scale, I subject the findings on party leader endorsements to a
series of robustness checks with non-parametric tests. While I do not include the results of that analysis here, both
the two-sample Wilcoxon-Mann-Whitney rank sum test and the two-sample Komolgorov-Smirnov test, which is
known to behave highly conservatively when used for discrete distributions (Conover 1972), replicate the results
from the parametric tests.
15
Figure 2: Effects of Party Leader Endorsements/Denouncements: Pooling across Performance
Dimension, TNA and ODM Sample
Note: Pooled sample with both TNA and ODM primary voters. The figure reports point estimates for the mean of
each treatment condition. The thick and thin lines represent 95 and 99% confidence intervals for the means. The
difference in means is derived from a standard two-tailed t-test. ***p<0.01, **p<0.05, *p<0.1. The results survive
the Benjamini-Hochberg FDR corrections at an FDR of 0.05.
dimension for both TNA and ODM primary voters.19 Figure 2 shows that the mean vote intention for the aspirant who was endorsed by the party leader (5.29) is on average 22% larger than
the mean vote intention of an aspirant who has neither been endorsed nor denounced by the
party leader (4.33). Movement from a 4 (neither likely nor unlikely to support) to a 5 (somewhat
likely to support) on the 7 point likert scale of vote intentions means that on average, the party
leader’s endorsement moves the otherwise ambivalent primary voter towards becoming favorable towards the aspirant. The mean evaluation of an aspirant who, on the other hand, has been
denounced by the party leader (3.69) is on average 15% smaller than that of an aspirant who has
neither been endorsed nor denounced (4.33). Movement from a 4 (neither likely nor unlikely
to support) to a 3 (somewhat unlikely to support) on the vote intention scale signifies that a
party leader’s denouncement will make an ambivalent voter to lean towards not supporting the
aspirant. The magnitude of endorsement/denouncement effects can be appreciated when we
consider the differential between an aspirant who has been endorsed by the party leader to an
aspirant who has been denounced by the party leader: moving from a denounced aspirant to
an endorsed aspirant leads to a 1.6 point (or 43%) increase in primary vote intention, meaning
that primary voters will switch from leaning against voting to leaning towards voting for the
19
The large dot represent point estimates for the mean of each condition, and the thick and thin lines represent
95% and 99% confidence intervals for the means. The difference in means across the categories, marked with the
horizontal brackets on the plot, are derived from standard two-tailed t-tests.
16
aspirant.
Figure 3: Effects of Party Leader Endorsements/Denouncements: Pooling across Performance
Dimension, Disaggregated by Party
(a) TNA (Incumbent) Party
(b) ODM (Opposition) Party
Note: Sample disaggregated by party. The figure reports point estimates for the mean of each treatment condition.
The thick and thin lines represent 95 and 99% confidence intervals for the means. The difference in means is derived from a standard two-tailed t-test. ***p<0.01, **p<0.05, *p<0.01. The results survive the Benjamini-Hochberg
FDR corrections at an FDR of 0.05.
These results are replicated when the sample is disaggregated by party: as seen in columns
(2), (3), (5), and (6) in Table 3 as well as Figure 3, both TNA and ODM primary voters adjusted their vote intentions for the aspirant when their party leaders, Uhuru Kenyatta and Raila
Odinga, endorsed or denounced the aspirant. Interestingly, President Uhuru Kenyatta’s assessment of the candidate seems to have a larger impact on his co-partisan (TNA) primary voters
than Raila Odinga’s opinion has over his co-partisans (ODM). While the effect of both leader’s
endorsement of the candidate has a large and statistically significant effect, the size of the effect of Kenyatta’s endorsement (1.13 on a 7 point scale) is around 30% larger than the effect of
Odinga’s endorsement (0.81 on a 7 point scale). Similarly, the size of the effect of Kenyatta’s
denouncement (0.81 on a 7 point scale) is larger than that of Odinga’s (0.46 on a 7 point scale).
This finding is contrary to qualitative accounts of both local political analysts20 and politicians21
themselves, who often suggested that Odinga’s influence over his co-partisans would be much
larger than Kenyatta’s. While it is unclear why this disparity in effect sizes exists, a possibility is
that it reflects the prolonged exposure of partisans to news coverage over the contentious na20
Transcript of interview with political journalist, Interview Subject 2016-EJ8, conducted on June 10, 2016
Transcript of interview with sitting members of parliament, Interview Subject 2015-PI24, conducted on April
24, 2015; Interview Subject 2015-PO27, conducted May 2, 2015
21
17
ture of ODM primaries in the run-up to the 2017 elections, in which allegations of favoritism
and cronyism were repeatedly played.
Are the endorsement effects outlined in the preceding paragraphs uniquely attributable to
the party leader and his influence over his co-partisans? Without additional evidence, it is difficult to confirm whether or not the treatment effect observed here is attributable to the party
leaders themselves or whether a generic endorsement from any other political or non-political
actor would have had similar effects on aspirant evaluations. By virtue of the setup of the experiment, however, it is possible to get some traction into this important point: in the design,
I also included a condition in which a locally-elected party elite (member of county assembly,
hereafter MCA) also endorses/denounces the aspirant. Given the growing importance of local
politicians such as MCAs in the context of Kenya (Arriola, Choi and Gichohi, 2016), comparing
the effect of party leader endorsements effects to MCA endorsement effects constitutes an important first hurdle. Figure B1 in Appendix B overlays the MCA endorsement/denouncement
conditions to Figure 2. As is clear from the results of the difference in means tests, the mean vote
intention of the party leader endorsement condition is around 0.52 points higher than the MCA
endorsement condition. Similar results hold for the denouncement conditions. Combined with
baseline comparisons in Table 3 and Figure 2, these findings suggest that party leaders have a
substantively large influence over primary voters and how they evaluate candidates.
Evaluation of Other Candidate Attributes: Potential Mechanisms
While the design of this experiment does not give us full inferential leverage over the potential
mechanisms that might be driving these findings, it does present us with some suggestive hints.
Figure 4 first presents the results of difference in means across the party leader endorsement/ denouncement conditions and the no endorsement control conditions for post-treatment
evaluations on various aspirant attributes: these were perceived loyalty to the party, loyalty to
the party leader, willingness to campaign for the party to take the presidency, ability to help the
constituency, likelihood of becoming the MP, and trustworthiness. Interestingly, aspirants with
party leader endorsements are favored over aspirants with no endorsement across all 6 of the
attribute evaluation survey items. The converse is true for aspirants who are denounced by the
party leader. The difference in means for aspirant evaluations across the six survey items are
always negative, meaning that aspirants who were denounced by the party leader were rated
poorly in comparison a candidate without an endorsement. While the largest differences are
observed in the perceived loyalty of the aspirant towards the party leader and the party, the
fact that there are statistically significant differences across all the survey items prevent us from
triangulating on which of the aspirant characteristics can be held accountable for the shift in
overall vote intention. However, an interpretation of these data is that, as predicted in the theoretical section of the paper, the party leader’s endorsement or denouncement had a profound
effect on the how respondents evaluate primary aspirants, across multiple dimensions including evaluations about the candidate’s anticipated job performance for which respondents were
provided with explicitly had countervailing information in the treatment.
18
Figure 4: Mechanisms - Party Leader Endorsement/Denouncement Effects
(a) Leader Endorsed - No Endorsement
(b) Leader Denounced - No Endorsement
Note: The thick and thin lines represent 95 and 99% confidence intervals for the means. The difference in means is
derived from a standard two-tailed t-test. Results of the analyses used to generate this figure is presented in Table
B1, Appendix B.
Can Leader Endorsement Effects Counteract the Effects of Aspirant Quality/Performance?
The design of the experiment also allowed me to manipulate the nature of the information on
aspirant quality/performance simultaneously with the endorsement dimensions. By doing so, I
can investigate the extent to which the aspirant’s prior performance (in this context, how much
he/she contributed to constituency service and local development) affects voter evaluation of
the primary aspirant. Figure 5 presents the results of the analysis, for which I pooled across
all the endorsement dimensions. The findings are largely similar to what existing literature has
found: primary voters reward good performance. The mean primary vote intention for a high
performance aspirant (4.88) is 0.8 points larger than that of a low performance aspirant. This
difference is statistically significant at the p<0.001 level, and is robust to other non-parametric
tests and survives the Benjamini-Hochberg adjustment for multiple testing.
How does the size of the performance effect compare with the effect of endorsements or
denouncements? In order to juxtapose the effect of party leader endorsements and denouncements with the effect of candidate quality / performance, I leverage the factorial design of the
experiment. The corresponding results are presented in Table 4. The basic intuition behind
the statistical tests presented in Table 4 is that if the magnitude of party leader endorsements
and denouncement are indeed large and significant, it should be able to offset the evaluation
gap induced by the difference in aspirant quality or performance. For example, if the impact
of party leader endorsements is large enough, there should be little to no observed difference
between a low performance aspirant who has been endorsed by the party leader and a high
19
Figure 5: Effects of Candidate Quality/Performance: Pooling across Endorsement Dimension,
TNA and ODM Sample
Note: Pooled sample with both TNA and ODM primary voters. The figure reports point estimates for the mean of
each treatment condition. The thick and thin lines represent 95 and 99% confidence intervals for the means. The
difference in means is derived from a standard two-tailed t-test. ***p<0.01, **p<0.05, *p<0.1. The results survive
the Benjamini-Hochberg FDR corrections at an FDR of 0.05.
performance aspirant. Conversely, if the impact of party leader denouncements is sufficiently
large, there should be little to no observable difference between the evaluation of a high performance aspirant who has been denounced by the party leader and a low performance candidate.
As the results presented in Table 4 suggest, the effects of party leader endorsements and
denouncements are large enough that they often offset the effect that aspirant quality/ performance has on aspirant evaluations. For example, Test 1 of Table 4 shows that there is no
statistically significant difference between a high performance aspirant without an endorsement and a low performance aspirant who has the endorsement of the party leader. Similarly,
Test 3 of Table 4 shows that there is only a marginally significant difference between a low performance aspirant without an endorsement and a high performance aspirant who has been
denounced by the party leader. In some cases, the effects of party leader endorsements and
denouncements overpower the combined effect of aspirant performance and the endorsement
or denouncement of a local politician: as Test 2 shows, the difference in the evaluation of a low
performance aspirant with a party leader endorsement is only marginally smaller than a high
performance aspirant with an MCA endorsement. Furthermore, any statistically significant differences except for test 4 disappear when I adjust the p-values for multiple testing using the
B-H correction.
20
Table 4: Party Leader Effects versus Aspirant Performance Effects
Test 1. T6 - T5
: Low performance, leader endorsed
- High performance, no endorsement
Test 2. T6 - T2
: Low performance, leader endorsed
- High performance, MCA endorsed
Test 3. T3 - T10
: High performance, leader denounced
- Low performance, no endorsement
Test 4. T3 - T8
: High performance, leader denounced
- Low performance, MCA denounced
a
b
Diff. in
means
Sig. w/FDR
correction
Rank sum test
(p-value)
K-S test
(p-value)
-0.046
No
0.972
1.000
-0.287*
No
0.054
0.253
0.299*
No
0.064
0.395
0.477**
Yes
0.005
0.013
Difference-in-means are assessed using a standard two-tailed t test. ***p<0.01, **p<0.05, *p<0.1. For
the multiple testing adjustment, I use the Benjamin-Hochberg FDR correction at an FDR of 0.05.
Tests included in this table are pre-registered in the analysis plan. The number following the alphabet T refers to a specific treatment condition, as labeled in Table 3.
Heterogeneous Treatment Effects
The results thus far have demonstrated that the effect of party leader endorsements and denouncements have a large impact on the vote intention of partisan primary voters, and that
the size of the effects are sometimes large enough to offset the effect of aspirant performance. I
have also found some suggestive evidence that the endorsement and denouncement effects are
likely mediated by the voter’s assessment regarding the aspirant’s attributes, including loyalty
towards the party, the party leader, and his anticipated job performance. In this section, I assess
whether the effect of party leader endorsements and denouncements are moderated by certain
respondent attributes and characteristics: are certain types of partisan primary voters likely to
respond more strongly to the opinion of their party leaders? For example, are partisans with a
stronger sense of linked fate with the party leader more inclined to listen to the opinion of the
party leader? Are the party leader’s coethnic partisans primarily responsible for the endorsement and denouncement effects? Are the endorsement effects moderated by the respondent’s
prior evaluation of the party leader’s job performance? Do low information voters in particular
privilege the word of the party leader to caste their vote in the party primaries?
21
Table 5: Heterogeneous Effects of Leader Endorsements/Denouncements
Leader Endorsement Effect
(1)
Linked Fatea
Linked Fate x Tr
(2)
(3)
(6)
(8)
-0.635***
(0.219)
0.294
(0.255)
-0.008
(0.075)
0.055
(0.104)
Lr. Approval x Tr
-0.037
(0.077)
0.033
(0.113)
0.944***
(0.104)
0.656***
(0.194)
0.955***
(0.106)
0.083
(0.273)
-0.102
(0.367)
0.963***
(0.110)
X
X
X
X
X
X
X
X
X
X
937
0.087
1002
0.045
Low Knowledged
Lo Knowledge x Tr
936
0.093
(7)
-0.147*
(0.084)
0.159
(0.116)
Leader Approvalc
Sample
N
R2
(5)
-0.501*
(0.200)
0.402*
(0.231)
Coethnic x Tr
Controls
Location FE
(4)
-0.090
(0.084)
0.253**
(0.105)
Coethnicb
Treatment
Leader Denouncement Effect
Pooled (TNA + ODM)
937
937
0.093
0.088
-0.663***
(0.113)
-0.885***
(0.216)
-0.654***
(0.114)
0.052
(0.271)
-0.072
(0.421)
-0.648***
(0.119)
X
X
X
X
X
X
Pooled (TNA + ODM)
1003
1003
0.050
0.042
1003
0.042
Robust standard errors in parentheses. ***p<0.01, **p<0.05, *p<0.1.
All variables except for dichotomous variables are standardized for the analyses.
a
Linked fate is measured using the question “Do you think what happens to your party leader will affect what
happens in your life? If yes, how much will it affect you?” Responses were recorded on a 4 point scale ranging
from 1 (None) to 4 (Yes, a lot).
b
A dichotomous variable indicating whether the respondent was a coethnic of the party leader.
c
Leader approval is measured using the question “Do you approve or disapprove of the way the party leader of
[insert party name here] is handling his job? Responses were recorded on a standard 7 point likert scale.
d
Coded based on an open-ended question asking the respondent to name the current MP of her constituency,
as well as the MP’s party affiliation. Low knowledge is a dichotomous variable that takes on a value of 1 when
the respondent failed to correctly identify both the name and the party affiliation of the MP.
In order to test how the effect of leader endorsements and denouncements are moderated,
I conduct a heterogeneous treatment effects analysis where I regress our main outcome against
four moderators measured at the individual level, the treatment indicators, and the interactions of the moderators and treatment indicators. The four moderators, as specified in the
pre-analysis plan are 1) level of linked fate with the party leader, 2) coethnicity with the party
leader, 3) job approval of the party leader, and 4) level of respondent political knowledge. I
also included a battery of respondent characteristics including gender, religion, ethnicity, and
a self-assessment of their living conditions and location fixed effects as controls. The specific
22
regression equation estimated was as follows:
Yi = β0 + β1 Mod er at or i + β2 Ti + β3 Mod er at or i × Ti + β4 X i + δ j + ²i
(1)
where Ti denotes the treatment status of the respondent, X i is a vector of individual-level
covariates measured pre-treatment, and δ j is a dummy for respondent location.
The results of the analyses are presented in Table 5. Columns (1) and (5) shows the results
for whether the treatment effect for party leader endorsements and denouncements are moderated by respondent’s perception of linked fate with the party leader. The evidence seems to
be asymmetrical: individuals who report higher levels of linked fate with the party leader respond more strongly to the endorsement treatment, as observed in the large positive sign of
the interaction term between linked fate and the treatment status indicator. The effect is statistically significant at p<0.05. The moderating effects of linked fate, however, are not observed
in relation to the denouncement treatment: the coefficient for the interaction term between
the linked fate measure and the treatment indicator is positive, but is not statistically significant at conventional levels. The same asymmetry is observed for the coethnicity with the party
leader. Whereas individuals who are coethnics of the party leader are more likely to be supportive of a primary aspirant who has been endorsed by the party leader (marginally significant at
p<0.10), no such moderating effects are observed for coethnicity regarding the denouncement
treatments. Neither the prior levels of approval for the party leader nor the level of the respondent’s political knowledge seem to be moderating the effect of party leader endorsements and
denouncements: none of the multiplicative terms, reported in columns (3), (4), (7), (8), meet
standard levels of statistical significance.
Overall, the results of the exploratory analyses adds partial credence to the idea that linked
fate and coethnicity between the party leader and respondents will moderate the effect of party
leader endorsements, providing some corroboration to the theoretical discussion presented in
section 2. While we find very little evidence of further moderating effects with regard to respondent evaluation of leaders and their level of political knowledge, this maybe due to the
limited variation on these characteristics within our sample: almost 90% of respondents have
a favorable evaluation of the party leader’s job performance, while only 10% of respondents incorrectly stated both the name and the party of their current MP. It is also worth highlighting
again that the analyses conducted here are exploratory, and any conclusions that can be drawn
are tentative.
6 Experiment 2: Conjoint Analysis
6.1 Experimental Design and Methods
The first experiment provides strong evidence that party leader endorsements and denouncements shape how partisan primary voters evaluate political aspirants. Does the effect of party
leader endorsements and denouncements hold in a multiple candidate framework in which
voters are presented with viable alternatives? For example, would a voter’s preference for a
leader endorsed aspirant hold when voters have an option to choose another aspirant who has
a strong record of local development? In the second experiment, I move beyond the single candidate framework and test the effect of party leader endorsement on the intended vote choice of
23
primary voters. I specifically employ conjoint analysis, which allows for the simultaneous estimation of multiple treatment components using a discrete choice task that mirrors the choice
that voters face in the ballot box in a typical election: one in which voters cast their vote for a
single candidate from a set of candidates that differ along multiple attributes and dimensions.
While only a handful of studies have examined the determinants of voter behavior specifically in the context of primary elections, I combine the insights from those studies with other
candidate attributes that have been found in prior research to influence vote choice in general
election settings across Africa. These attributes and attribute levels are presented in Table 6.
Table 6: Conjoint Analysis - Candidate Attributes and Attribute Levels
Candidate Attributes
Attribute Levels
Current occuption
Member of parliament (MP)
Business owner
Professor at a university
School teacher
Kikuyu/Kalenjin/Kamba/Luo/Luhya
Male/Female
Cabinet Minister
Deputy Minister
None
Largest donation to the school renovation project
Did not donate to the school renovation project
Largest donation for new health clinics
Did not donate to new health clinics
Provided bursaries for 150 children in the constituency
Did not provide bursaries for children in the constituency
Paid the hospital fee for 150 sick people
Did not pay the hospital fee for sick people
None (unknown)
Convicted of corruption for handing out cash to voters
Under investigation for embezzling funds for personal use
No record of corruption
publicly stated that he strongly supports the candidate
publicly stated that he does not support the candidate
has not expressed his opinion about the candidate
Ethnic group
Gender
Previous government appointments
Contribution to local development
Record on corruption
Party leader’s position
Profiles of fictitious aspirants for the TNA and ODM primaries were randomly generated using the attributes and attribute levels in Table 6. Though the total number of possible combination of attribute values is much larger than what would be actually observed in reality, the random assignment of attribute values22 guarantees that profiles with a certain attribute-attribute
22
For the candidate ethnic group attribute, we deviate from convention and do not assign with equal probability:
instead, we use the population proportion of the ethnic groups based on the most recent census data on ethnic
group distributions in Kenya. This is to mitigate concerns raised by enumerators and respondents during piloting
that questioned the frequency with which candidate profiles with minority ethnic group membership were being
generated. For constituencies in Nakuru county, the ethnicity of the candidate were assigned according to the
following probability: Kikuyu 61%, Kalenin 15%, Kamba 6%, Luo 10%, Luhya 8%. For constituencies in Kisumu
county, the probability was as follows: Kikuyu 1%, Kalenjin 2%, Kamba 1%, Luo 90%, Luhya 6%. These probabilities
are accounted for in the analyses of the conjoint data.
24
level combination will have the same distribution for all other attributes on average as compared to profiles with the same attribute but a different attribute value level, allowing for a simple comparison means. Following a pre-treatment survey measuring standard demographic
information, the experimental respondents were presented with two profiles and asked “Which
of these two candidates would you prefer to vote for in the TNA/ODM party primaries?”.23 Per
common practice in conjoint experiments, this process was repeated 3 times per respondent,
for a total of 7,176 aspirant profile pairs and 14,352 individual aspirant profiles rated across the
two study locations.
6.2 Results
The quantity to be estimated is the average marginal component effect (AMCE). I use the fully
non-parametric linear regression estimator presented in Hainmueller, Hopkins and Yamamoto
(2013), and cluster the standard errors derived from the estimation at the respondent level.24 I
also estimate the conditional AMCEs to detect heterogeneity in treatment effects across the two
different party samples.
Figure 6 and Figure A1 and A2 in Appendix A report the main findings of the conjoint analysis. The dots and lines in the plots represent the point estimates and 95% confidence intervals
for the AMCEs of each attribute value on the probability that respondents chose the aspirant
in the choice task. Rows without any estimates represent the reference categories within each
attribute. The regression model that the plot was generated from is included in the online supplementary index.
In support of the main hypothesis, likely primary voters of the two parties seem to strongly
prefer aspirants who have been endorsed by the party leader: as seen in Figure 6, compared to
an aspirant for whom the party leader has not expressed an opinion, endorsed aspirants are 8.1
percentage points (SE=0.9) more likely to be preferred in the party primaries. In comparison to
an aspirant that has been denounced, an endorsed aspirant is more than 10 percentage points
more likely to be chosen as the preferred candidate. This finding retains statistical significance
after correcting for multiple testing using the Benjamini-Hochberg FDR correction at an alpha
level of 0.05. The results from the samples disaggregated by political party tells a similar story:
as presented in Figure A1 and A2 in Appendix A, endorsed aspirants are around 8 percentage
points more likely to be preferred as the party candidate in comparison to the baseline category
of aspirants for whom the party has not expressed his opinion. The size of these effects across
the two parties are remarkably similar (TNA - 8.4 percentage points, ODM - 7.8 percentage
points), providing assurance that the effect of party leader endorsements in the pooled sample
23
To minimize the possibility that respondents privilege the first attribute they encounter in the party profiles
(primacy effects) to guide their choice, I randomize the order of the attribute presented across respondents, but
hold the order constant within the respondent.
24
For example, the estimation of the AMCEs for party leader endorsement attribute is conducted by running the
following regression:
choi ce i j k = θ1 + θ2 [suppor t i j k = yes] + θ3 [suppor t i j k = no] + ²i j k
(2)
where choiceijk is the choice outcome, and [supportijk = yes], [supportijk = no] are dummy variables coded 1 if the
respondents are assigned these attribute levels. The reference category is the candidate where the party leader has
not expressed an opinion about the candidate.
25
Figure 6: Effects of aspirant attributes on probability of being preferred in primary elections:
Pooled sample - TNA and ODM
Note: Pooled sample with both TNA and ODM primary voters. Estimates are based on the benchmark OLS model
with standard errors clustered at the respondent level. The dots and lines represent point estimates for the AMCEs
while the bars represent 95% confidence intervals. Rows without any estimates represent the reference categories
within each attribute.
are not being driven by any one of the two parties included).
Although I detect statistically significant effects for party leader endorsements, a few caveats
are in order: first of all, the coefficient for the party leader endorsement attributes, while substantively large, is smaller than some of the other attributes included in the experimental design. For example, the AMCEs for the positive performance attribute levels in the “contribution to local development” attribute as well as the “record on corruption” attribute are much
larger than the coefficient for party leader endorsements (the coefficient for these attribute levels range from 17-27 percentage points). There are two potential interpretation of this large size
differential: first, it maybe that the large effects for these candidate quality attributes reflect the
importance African voters place on the performance of their politicians (Conroy-Krutz, 2013;
Carlson, 2015; Adida et al., 2016). A second explanation might be that the relatively smaller effects for the endorsement attributes are the result of the non-specificity of the wording included
in the conjoint design. Whereas the wording for the candidate quality attribute were generally
more specific (invoking specific initiative and projects that the aspirant contributed to), the attribute for party leader endorsements were less specific in that it did not invoke the name of the
party leader, and did not describe in any detail the context of the endorsement. Anecdotal ac26
counts by survey enumerators suggest that many respondents asked follow up questions about
the endorsement attribute, including inquiries about when and where the endorsement was
given, and the overall nature of the relationship between the candidate and the party leader.
Second, it is also worth noting that the effects of party leader endorsement and denouncements might be asymmetrical: consistently across the pooled sample and the disaggregated
individual party samples, the size of the denouncement attribute level is significantly smaller
than the endorsement attribute level, and is only marginally statistically significant at p<0.1.
While the theoretical framework laid out in the previous section does not provide an a priori
reason to expect this asymmetry, it might be reflective of the difference in how respond to positive versus negative information, and how that interacts with baseline expectations of politician
behavior (Adida et al., 2016).25
7 Discussion and Conclusion
In this paper, I developed an argument regarding why and how party leaders in new democracies may attempt to retain their grip over candidate selection processes even when intraparty institutional reforms have removed the de jure authority for them to do so. I argued that
contrary to the existing literature, party leaders are incentivized to influence the outcome of
bottom-up processes of candidate selection such as primaries because they have no guarantee that the grassroots elections will select candidates who are loyal and compliant with them.
Given that divided loyalties among elected party elites have the potential to undermine the
party leader’s position within the party, party leaders will resort to strategies that attempt to directly induce to primary voters to “vote with the party leader”. I also argued that one of the most
common and potentially influential strategies are endorsements and denouncements issued
by the party leader to her followers informing them of their preferences over a specific aspirant.
The strong sense of linked fate forged between party leaders and partisans endows heuristic
value to the endorsement or denouncement, which voters rely on to make evaluations about
aspirants during primary elections.
I test the implication of the theory that party leader endorsements will have a strong persuasive effect for candidate evaluations and primary vote choice. Using two separate experimental
designs implemented on a large sample of 2392 likely primary voters of two major incumbent
and opposition political parties in Kenya, I show consistently that party leader endorsements
and denouncements have a strong effect on primary vote intention and candidate evaluation.
The effects are substantively large, and often counteracts the effect of other relevant information about the aspirant. They are also robust to corrections for multiple testing, a common
problem for analysis of experimental data.
While the study has provided some interesting insights regarding the development of parties in new democracies, I raise some potential for follow-up projects or extensions. First of
all, while beyond the scope of the paper, a detailed descriptive analysis of these party leader
25
It is interesting that the asymmetry is observed for the aspirant performance attributes. In comparison to the
baseline category where no information was given regarding the performance of the aspirant, the AMCEs for the
attribute levels with negative information on the aspirant’s performance were not statistically distinguishable from
zero. This also lends suggestive evidence in support of the idea that respondents are reacting to and processing
positive versus negative information in different ways.
27
endorsements and denouncements - for example, under what conditions these strategies are
employed and what specific forms these take - was wanting throughout the paper, and is likely
to be an important contribution to the study of candidate selection and party politics more
broadly. Second, while special attention was paid to enhance the reality of the radio news treatments in the second experiment, the inherently artificial nature of being aware that one is being
studied - a common problem for much of the survey-based research - possibly warrants an extension of this project into an observational or field experimental setting where the effects of
real world endorsements are analyzed vis-a-vis real-world electoral data.
28
Appendix A: Auxiliary Figures for Experiment 1
Figure A1: Effects of aspirant attributes on probability of being preferred in primary elections
Note: Results from ODM (opposition) primary voters. Estimates are based on the benchmark OLS model with
standard errors clustered at the respondent level. The dots represent point estimates for the AMCEs while the bars
represent 95% confidence intervals. Rows without any estimates represent the reference categories within each
attribute.
29
Figure A2: Effects of aspirant attributes on probability of being preferred in primary elections
Note: Results from TNA (incumbent) primary voters. Estimates are based on the benchmark OLS model with
standard errors clustered at the respondent level. The dots represent point estimates for the AMCEs while the bars
represent 95% confidence intervals. Rows without any estimates represent the reference categories within each
attribute.
30
Appendix B: Auxiliary Figures for Experiment 2
Figure B1: Effects of Party Leader Endorsements/Denouncements: Pooling across Performance
Dimension
Note: Pooled sample with both TNA and ODM primary voters. The figure reports point estimates for the mean of
each treatment condition. The thick and thin lines represent 95 and 99% confidence intervals for the means. The
difference in means is derived from a standard two-tailed t-test. ***p<0.01, **p<0.05, *p<0.1. The results survive
the Benjamini-Hochberg FDR corrections at an FDR of 0.05.
31
Table B1: The Effect of Party Leader Endorsements / Denouncements
A. Party Leader Endorsement Effects
Primary
Vote
Loyal
to Party
Loyal
to Leader
Campaign
for Party
Help
Const.
Trustworthiness
(1)
(2)
(3)
(4)
(5)
(6)
0.823
(0.089)
1.046
(0.084)
0.867
(0.092)
0.644
(0.099)
0.652
(0.095)
936
0.083
923
0.140
921
0.087
924
0.043
929
0.072
ITTa 0.953
SEb (0.105)
N
R2
937
0.080
B. Party Leader Denouncement Effects
Primary
Vote
Loyal
to Party
Loyal
to Leader
Campaign
for Party
Help
Const.
Trustworthiness
(1)
(2)
(3)
(4)
(5)
(6)
-1.089
(0.100)
-1.471
(0.102)
-1.067
(0.109)
-0.562
(0.111)
-0.854
(0.102)
1001
0.105
985
0.174
989
0.087
987
0.025
995
0.065
ITTa -0.647
SEb (0.114)
N
R2
a
b
c
1003
0.031
Estimated Intention-to-Treat Effects (ITTs) of the party leader endorsement / denouncement treatments on the main outcome (vote intention in primary elections)
and other candidate evaluations, pooling across performance dimensions. ATEs
are estimated against pure controls in which no endorsement information was provided.
Robust standard errors (SEs) from linear regression analysis.
All results survive the Benjamini-Hochberg FDR correction at an FDR of 0.05
32
Appendix C: News Script for Experiment 2
News script for Experimental Treatment
English script
Anchor: This is the news in Brief from KRN. I am Beatrice [Okelo/Njoroge]. Today, aspiring candidate for Member of Parliament, John [Oduor/Mwangi], addressed a gathering of constituents
to officially announce his intention to seek the [ODM/TNA] nomination for the 2017 elections.
During the rally, he spoke of his political qualifications and urged voters to support him during
the [ODM/TNA] primaries scheduled for early next year.
Candidate: I am a proud member of this community and have served this community for many
years. But our current leaders have repeatedly failed to deliver on their promises and we have
had enough. Today, I am announcing my intention to run for MP on an [ODM/TNA] party
ticket. I ask party members and voters of [Kisumu/Nakuru] to support my candidacy in the
[ODM/TNA] primaries as well as the general elections in 2017.
Anchor: By throwing his name into the mix, Mr. [Oduor/Mwangi] enters an already crowded
field of candidates for the [ODM/TNA] ticket. Many see the [ODM/TNA]] ticket as guaranteeing
the MP seat in an area dominated by [TNA/ODM].
Performance Dimension
(1) Candidate Performance High: Mr. [Oduor/Mwangi] is well-known for his strong record of involvement in the constituency’s development initiatives, including his major donations to the
school and classroom renovation initiative as well as his financial assistance for constituents
who cannot afford to pay for medical bills.
(2) Candidate Performance Low: Mr. [Oduor/Mwangi], is a newcomer to the political scene,
with only a limited record of involvement in the constituency’s community-driven development initiatives, to which many constituents expect aspiring politicians to make significant donations.
Endorsement Dimension
(1) Leader Endorsement: Mr. [Oduor/Mwangi]’s campaign is hopeful that primary voters will
choose him as the [ODM/TNA] candidate, given Mr. [Oduor/Mwangi]’s strong relationship
with [ODM party leader Raila Odinga/TNA party leader Uhuru Kenyatta]. [Raila Odinga/Uhuru
Kenyatta] is known to be strongly supportive of Mr. [Oduor/Mwangi]’s candidacy and his dedication to [ODM/TNA].
(2) MCA Endorsement: Mr. [Oduor/Mwangi]’s campaign is hopeful that primary voters will
choose him as the [ODM/TNA] candidate, given Mr. [Oduor/Mwangi]’s strong relationship with
local [ODM MCAs/TNA MCAs]. Many local MCAs are known to be strongly supportive of Mr.
[Oduor/Mwangi]’s candidacy, and his dedication to [ODM/TNA].
33
(3) Leader Denouncement: Mr. [Oduor/Mwangi]’s campaign dismissed concerns that he has
had a falling out with party leader [Raila Odinga/Uhuru Kenyatta]. Mr. [Oduor/Mwangi] was
noticeably missing from events attended by [Raila Odinga/Uhuru Kenyatta] during his recent
visit to [Kisumu/Nakuru]. Mr. [Odinga/Kenyatta] is known to be highly skeptical of Mr. [Oduor/
Mwangi]’s candidacy and his dedication to [ODM/TNA].
(4) MCA Denouncement: Mr. [Oduor/Mwangi]’s campaign dismissed concerns that he has had
a falling out with local [ODM MCAs/TNA MCAs]. Mr. [Oduor/Mwangi] was noticeably missing
from [ODM/ TNA] county party events. Many [ODM MCAs/TNA MCAs] are known to be highly
skeptical of Mr. [Oduor/Mwangi]’s candidacy and his dedication to [ODM/TNA].
(5) No Endorsement/Denouncement: Mr. [Oduor/Mwangi]’s campaign commented that while
the [ODM/TNA] primaries are very competitive, their candidate?s credentials will be most appealing to the voters.
Anchor: The early announcement of Mr. [Oduor/Mwangi] for the [ODM/TNA] party ticket
highlights how competitive the party primaries are expected to be.
34
References
Adida, Claire, Jessica Gottlieb, Eric Kramon and Gwyneth McClendon. 2016. “How Coethnicity Moderates the Effect of Information On Voting Behavior: Experimental Evidence from
Benin.”.
Adida, Claire L. 2015. “Do African Voters Favor Co-Ethnics? A New Identification Strategy.”
Journal of Experimental Political Science 2(1):1–11.
Arceneaux, Kevin. 2008. “Can partisan cues diminish democratic accountability?” Political Behavior 30(2):139–160.
Arriola, Leonardo R. 2012. Multi-Ethnic Coalitions in Africa: Business Financing of Opposition
Election Campaigns. New York: Cambridge University Press.
Arriola, Leonardo R., Donghyun Danny Choi and Matthew K. Gichohi. 2016. “Political Endorsements and Cross-Ethnic Voting.” Working Paper .
Baldwin, Kate. 2013. “Why vote with the chief? Political connections and public goods provision
in Zambia.” American Journal of Political Science 57(4):794–809.
Boudreau, Cheryl and Scott A MacKenzie. 2014. “Informing the electorate? How party cues
and policy information affect public opinion about initiatives.” American Journal of Political
Science 58(1):48–62.
Carey, John M and John Polga-Hecimovich. 2006. “Primary elections and candidate strength in
Latin America.” Journal of Politics 68(3):530–543.
Carlson, Elizabeth. 2015. “Ethnic voting and accountability in Africa: A choice experiment in
Uganda.” World Politics 67(02):353–385.
CGD, Center for Global Development. 2005. “Money and Politics: The Case of Party Nominations in Kenya.” NDI Policy Paper pp. 1–65.
Chandra, Kanchan. 2007. Why ethnic parties succeed: Patronage and ethnic head counts in India. Cambridge University Press.
Cheeseman, Nic. 2008. “The Kenyan elections of 2007: an introduction.” Journal of Eastern
African Studies 2(2):166–184.
Cobb, Rachel. 2006. “Who Nominates? The Role of Parties in Gubernatorial Nominations in Six
States.” Paper Presented at the Annual Meeting of the American Political Science Association .
Cohen, Marty, David Karol, Hans Noel and John Zaller. 2009. The party decides: Presidential
nominations before and after reform. University of Chicago Press.
Conroy-Krutz, Jeffrey. 2013. “Information and ethnic politics in Africa.” British Journal of Political Science 43(02):345–373.
35
Dawson, Michael C. 1994. Behind the mule: Race and class in African-American politics. Princeton University Press.
Duverger, Maurice. 1959. Political parties: Their organization and activity in the modern state.
Methuen.
Elischer, Sebastian. 2013. Political parties in Africa: Ethnicity and party formation. Cambridge
University Press.
Ferree, Karen E. 2010. Framing the Race in South Africa: The Political Origins of Racial Census
Elections. Cambridge University Press.
Field, Bonnie N and Peter M Siavelis. 2008. “Candidate Selection Procedures in Transitional
Polities A Research Note.” Party Politics 14(5):620–639.
Gichohi, Matthew K. 2016. “Ethnic Groupness and Party Competition.” Working Paper .
Hainmueller, Jens, Daniel J Hopkins and Teppei Yamamoto. 2013. “Causal inference in conjoint analysis: Understanding multidimensional choices via stated preference experiments.”
Political Analysis p. mpt024.
Harding, Robin. 2015. “Attribution and Accountability: Voting for Roads in Ghana.” World Politics 67(04):656–689.
Hazan, Reuven Y and Gideon Rahat. 2010. Democracy within parties: Candidate selection methods and their political consequences. Oxford University Press.
Hopkin, Jonathan. 2001. “Bringing the members back in? Democratizing candidate selection in
Britain and Spain.” Party politics 7(3):343–361.
Hyden, Goran and Colin Leys. 1972. “Elections and politics in single-party systems: the case of
Kenya and Tanzania.” British Journal of Political Science 2(04):389–420.
Ichino, Nahomi and Noah L. Nathan. 2012. “Primaries on Demand? Intra-Party Politics and
Nominations in Ghana.” British Journal of Political Science 42(04):769–791.
Ichino, Nahomi and Noah L. Nathan. 2013.
“Do Primaries Improve Electoral Performance? Clientelism and Intra-Party Conflict in Ghana.” American Journal of Political Science
57(2):428âĂŞ441.
Ichino, Nahomi and Noah L Nathan. 2016. “Who Gets on the Ballot: Money, Ethnicity, and
Candidate Entry in Ghana.” Working Paper .
Izama, Melinda Platas and Pia Raffler. 2016. “Meet the Candidates: Information and Accountability in Primaries and General Elections.” Working Paper .
Katz, Richard S. 2001. “The problem of candidate selection and models of party democracy.”
Party politics 7(3):277–296.
36
Khadiagala, Gilbert M. 2010. “Political movements and coalition politics in Kenya: entrenching
ethnicity.” South African Journal of International Affairs 17(1):65–84.
Koter, Dominika. 2013. “King makers: Local leaders and ethnic politics in Africa.” World Politics
65(02):187–232.
Lau, Richard R and David P Redlawsk. 2001. “Advantages and disadvantages of cognitive heuristics in political decision making.” American Journal of Political Science pp. 951–971.
Long, JD and B Hoffman. 2013. “Party attributes, performance, and voting in Africa.” Comparative Politics 45(1):127–146.
Lupia, Arthur. 1994. “Shortcuts versus encyclopedias: Information and voting behavior in California insurance reform elections.” American Political Science Review pp. 63–76.
Lynch, Gabrielle. 2014. “Electing the ?alliance of the accused?: the success of the Jubilee Alliance
in Kenya’s Rift Valley.” Journal of Eastern African Studies 8(1):93–114.
Madrid, Raúl L. 2012. The rise of ethnic politics in Latin America. Cambridge University Press.
Masket, Seth. 2009. No middle ground: How informal party organizations control nominations
and polarize legislatures. University of Michigan Press.
McCauley, John. 2014. “The Political Mobilization of Ethnic and Religious Identities in Africa.”
American Political Science Review 108(04):801–816.
McClendon, Gwyneth and Rachel Beatty Riedl. 2015. “Religion as a Stimulant of Political Participation: Experimental Evidence from Nairobi, Kenya.” The Journal of Politics 77(4):1045–1057.
Norris, Pippa. 1997. Passages to power: Legislative recruitment in advanced democracies. Cambridge University Press.
Öhman, Magnus. 2004. “The heart and soul of the party: Candidate selection in Ghana and
Africa.”.
Ostrogorsky, Moisey. 1902. Democracy and the organization of political parties. Lightning
Source Incorporated.
Pennings, Paul and Reuven Y Hazan. 2001. “Democratizing candidate selection causes and
consequences.” Party Politics 7(3):267–275.
Siavelis, Peter M and Scott Morgenstern. 2012. Pathways to power: Political recruitment and
candidate selection in Latin America. Penn State Press.
Sniderman, Paul M and Edward H Stiglitz. 2012. The reputational premium: A theory of party
identification and policy reasoning. Princeton University Press.
Van de Walle, Nicolas. 2003. “Presidentialism and clientelism in Africas emerging party systems.” The Journal of Modern African Studies 41(2):297–321.
37
Van de Walle, Nicolas. 2007. Meet the New Boss, Same as the Old Boss? The Evolution of Political
Clientelism in Africa. In Patrons, Clients, and Policies: Patterns of Democratic Accountability
and Political Competition. Cambridge University Press.
Weghorst, Keith R and Staffan I Lindberg. 2013. “What drives the swing voter in Africa?” American Journal of Political Science 57(3):717–734.
Widner, Jennifer A. 1993. The rise of a party-state in Kenya: From" Harambee!" to" Nyayo!". Univ
of California Press.
Willis, Justin and Ngala Chome. 2014. “Marginalization and political participation on the Kenya
coast: the 2013 elections.” Journal of Eastern African Studies 8(1):115–134.
38