Introduction

Why is voting behaviour markedly linked to ethnic identity in many of Africa’s new democracies? A predominant explanation provided by the instrumental voting theory is that ethnic voting behaviour is driven by voters’ rationalisations that candidates’ ethnicity signals expected performance (Bratton et al., 2012). Coethnic candidates are thought to understand coethnic voters’ preferences better, and these candidates, if elected, are expected to provide better access to public goods than non-coethnic candidates (Carlson, 2016). Some observational studies support these predictions (Bratton et al., 2012; Hoffman & Long, 2013). These studies, however, offer self-rationalisations of voting behaviour and may not reveal the reasons for the ethnic voting pattern observed. Recent experimental studies have instead tried to unravel the role of politicians’ ethnicity and performance by asking voters to choose between hypothetical candidates with varying characteristics (e.g. Carlson, 2015; Conroy-Krutz, 2013; Gutiérrez-Romero & LeBas, 2020). Although this literature has advanced our understanding of political behaviour, experiments simulating imaginary elections do not fully reflect the type of candidates that voters face in real elections. Albeit there are also experimental studies presenting real candidates to voters (e.g. Adida, 2015; Wantchekon, 2003), in these studies, voters rarely have concrete evidence on issues that matters to them, such as candidate performance or evidence of corruption, limiting our understanding of what is driving voting.

This article presents experimental evidence that challenges the predominant views on the motivations behind ethnic voting. Our experiment, embedded in a nationally representative survey in Kenya, shared with voters the official spending record of their Member of Parliament (MP) and the result of independent audits that probed for corruption. We focus on how MPs managed the Constituency Development Fund (CDF), the country’s main poverty reduction programme for which each MP receives about two million dollars per year and decides, in consultation with residents, which types of public goods to finance. In the experiment, we showed voters evidence of their actual MP’s performance while randomising three cues: (1) a reminder that the politician was a serving MP, (2) the MP’s political party affiliation, and (3) the MP’s full name, which directly conveys her/his ethnicity.

This article offers two important contributions to the literature. First, the article shows that voters, regardless of ethnicity or partisanship, have an inaccurate perception of how their MP spent public funds in their constituency. Second, showing voters that the distribution of public goods made by their MP does not match their preferred spending and prompting them to view their MP as partisan, reduces their approval and likelihood to vote for the MP. Voters also electorally punish partisan MPs for corruption whether copartisan or non-copartisan. The same pattern does not hold for ethnicity. In contrast to instrumentalist expectations, priming respondents to view their MP in ethnic terms does not impact their assessment of the politician’s performance or corruption. It even boosts the probability that the coethnic MP will be re-elected, escaping the consequences of poor governance. This preference for coethnic candidates supports the identity and expressive voting theories which suggest voters prioritise group loyalties (Horowitz, 1985).

Literature Review

Ethnicity and Partisan Shortcuts

The reasons why ethnic voting shapes voting behaviour in Africa remain widely debated (Hoffman & Long, 2013; Horowitz, 1985). An early explanation, provided by the social identity theory, argues that people believe that members of their group are the best to lead them because identity is internalised as an aspect of self-esteem, and psychological pressures make them evaluate out-group members more harshly (Tajfel & Turner, 2004). Building on these sociopsychological explanations, the expressive theory suggests that self-identity not only serves as a source of people’s self-esteem. In ethnically divided societies, voting is not a rational choice but a way for people to express their non-negotiable allegiance to their group (Horowitz, 1985).

For the identity and expressive theories, partisanship, like ethnicity, provides another identity, which can lead to group favouritism and exaggerate negative traits of non-copartisans. Even though African political parties are young, often transitory, and without strong ideology, they serve as a cue of commitments made to specific constituents.Partisanship in some democracies might be significant and determined apart from other social identities. In others, partisanship and ethnic identities are strongly correlated (Ichino & Nathan, 2013). In Kenya, for instance, no single ethnic group is big enough to win elections on its own. Hence ethnic groups form alliances as a convenience vehicle to compete in elections, but such collaborations are often internally fragile and short-lived (Horowitz, 1985). The regular assembling and competition among these alliances explain why voting patterns in Kenya, like other similar countries, resemble mere “ethnic arithmetic” (Elischer, 2008).Footnote 1 Despite constitutional reforms in 2010, such as more power devolution and an absolute majority needed to win presidential elections, Kenyan parties remain weak and a vehicle to mobilise ethnic groups (Cheeseman et al., 2019).

A more recent and prominent explanation of identity voting is offered by the instrumentalist theory. For this theory, candidates’ identity, such as ethnicity, is used by voters as a shortcut to forecast candidates’ performance (Conroy-Krutz, 2013). “[V]oters prefer coethnic politicians because they expect coethnics to provide better future economic and political goods” and have “better knowledge of what goods their coethnic want” (Carlson, 2015, 354, 356). Partisanship, like ethnic identity, serves voters as a relevant cue of which candidate is more likely to reward their group with preferential policies and better goods (Wantchekon, 2003). That is, voters support the candidates with whom they share an identity because those candidates are expected to deliver public goods that more closely aligned with their preferences than those offered by candidates’ out-groups.

Albeit the identity/expressive and instrumental theories follow different logics, based on this discussion we expect:

Hypothesis 1

In the absence of other information, voters will assume that coethnic (H1a) or copartisan (H1b) politicians’ spending on public goods more closely aligns with their preferred funding allocation than politicians with whom they have no affective ties.

Performance

Voting theories differ more substantially in predicting whether voters’ choices will be affected if they receive information about politicians’ performance. For the identity and expressive theories, ethnic voting is purely identity driven. Thus, the voting choice will not be expected to change if voters become aware of politicians’ performance. Because identity makes people process information in a way that confirms their view, they will embrace information that reflects well on a politician from their group, and they will disregard any negative information (Tajfel & Turner, 2004). Voters will also rate rival candidates more negatively than candidates from their group. Nonetheless, for the identity/expressive theory, there can be instances where voters could consider politicians’ bad performance. If group evaluations render unfavourable comparisons for people’s self-identity, “individuals will strive either to leave their existing group and join some more positively distinct group and/or make their existing group more positively distinct” (Tajfel & Turner, 2004, p. 16). Voters could penalise politicians for bad performance, particularly if sharing a weaker identity attachment, as partisanship, and if there are alternative candidates better representing their group interests. Some empirical studies have found support for these predictions (Gutiérrez-Romero, 2013). For instance, in a survey experiment in Uganda, Carlson (2015) shows that when voters are offered a choice between two hypothetical candidates, a coethnic and non-coethnic, positive past performance cues increase the support for the politician; but only when the politician is a coethnic of the voter. Non-coethnic candidates cannot compensate by demonstrating good performance. Nonetheless, several other African studies have also shown that performance evaluations can influence voting choice in real elections, albeit modestly (Bratton et al., 2012; Hoffman & Long, 2013; Long & Gibson, 2015).

The evidence that voters consider politicians’ performance is stronger in African legislative elections (Barkan, 2009). The legislature helps citizens achieve representative governance and development. In Africa, MPs are often given local development funds to provide club public goods such as building schools, clinics, and roads in their constituencies. Where political clientelism is prevalent, MPs may also give jobs and cash to their clients (Lindberg, 2010). Thus, MPs’ performance is visible and measurable through goods and services provided. This visibility may explain why less than 50% of African legislative seat holders get re-elected (Barkan, 2009). In Ghana, for instance, over half of voters have switched parties based on their MP’s performance measured by goods delivered to the constituency and their legislating skills (Weghorst & Lindberg, 2013). In Kenya, the disputed 2007 elections that sparked turmoil and ethnic violence may have shown citizens that pitting one group against another can plunge the country into chaos when political competition is dominated by ethnic considerations (Cheeseman et al., 2014). But, even before that disputed election, most Kenyans stated that ethnicity did not guide their vote. In a nationally representative study, most voters (99%) said they chose their MP in 2007 based on their care for their community, honesty, and how they managed the constituency development fund (Gutiérrez-Romero, 2013). In these legislative elections, competition is so fierce that less than 50% of MPs contend for re-election.

The instrumental theory offers an alternative explanation of how voters weigh politicians’ performance. For this theory, even if ethnicity or partisanship is used initially to forecast the candidates’ performance, rational voters will update their expectations once more reliable information becomes available (Conroy-Krutz, 2013). Regardless of politicians’ identity, voters will reward well-performing politicians by being more likely to vote for them and punish poor performance by being more likely to switch support to opposition candidates.

An issue with instrumentalist explanations is that politicians’ actions and attribution are not easily observable. This is why incumbent governments in Africa perhaps provide more public investment and better services to their constituents in the form of roads and health services, for instance, than to opposition supporters (Burgess et al., 2015; Franck & Rainer, 2012).Footnote 2 These goods are cost-effective because voters can easily observe them and help evaluate politicians’ performance (Ichino & Nathan, 2013). Increasing transparency has thus been argued as a potential solution to voters’ asymmetries of information (Besley, 2005). Still, it is unclear how voters react to more transparency concerning their politicians since pre-existing beliefs could bias how voters process information.

Given the complexity of public spending, voters likely have asymmetries of information. Thus, we would expect that voters have inaccurate perceptions about how, in practice, their MP distributes expenditure in their constituency (on education, roads, electricity, etc.). We also hypothesise that voters would still be more likely to vote for their coethnic MP even if they received information that their coethnics did not spend funds closely aligned with their preferred spending. As the identity theory argues, ethnicity evokes group loyalty and is unlikely to trigger punishment because of a priori assumptions about coethnics. In contrast, we expect voters will be more likely to penalise deviations from their preferred spending and reduce their approval if they are prompted to think about their MPs as partisans. Given that voters often have alternative parties expressing their interests in elections, partisanship cue is more likely directly tied to performance expectations. Based on this, we hypothesise:

Hypothesis 2a

After learning that the spending on public goods made by their MP does not align with their preferred spending allocation, voters will be more likely to re-elect their coethnic MP when prompted to think about politicians in terms of their ethnicity.

Hypothesis 2b

After learning that the spending on public goods made by their MP does not align with their preferred spending allocation, voters will not be more likely to re-elect their MP when prompted to think about politicians as partisans.

Corruption

A concern with ethnic voting is that people could fail to sanction politicians not only for poor performance but also for criminal behaviour, such as corruption, leading to underdevelopment, democratic instability, ethnic conflict, and discriminatory provision of public goods (Banerjee & Pande, 2007; Easterly & Levine, 1997). A prominent explanation for why voters fail to punish corrupt politicians is that in-group loyalties are prioritised (De Vries & Solaz, 2017). According to the identity/expressive theories, even if voters disapprove of corruption, voters need to weight candidates’ negative evaluations against other internal psychological pressures (Horowitz, 1985). Voters thus might place greater weight on benefits they expect from a candidate, such as in-group status, than they do on that individual’s wrongdoing. Some experiments support this prediction. For instance, in Benin, Adida (2015) shows that the coethnic voters of Yayi Boni, the then incumbent President, were more likely to vote for him than non-coethnics, despite being told that detractors viewed his administration as corrupt and providing unsatisfactory growth.

For the identity theory, there may be instances where voters electorally punish their politicians for corruption. If comparisons to other groups lower voters’ self-esteem, voters can prioritise other identities, such as ethnicity, and punish politicians for corruption, particularly those with weaker self-identity attachments as copartisans. Banerjee and Pande (2007) formalise these predictions in a theoretical model (which finds empirical support in India) where voters prefer higher-quality candidates but differ in their policy preferences, reflecting differences in group identity. When a few parties represent groups’ interests, such as ethnic interests, voters will sacrifice the quality of candidates to elect a candidate that can deliver their own-group policy preference. In contrast, with many parties representing the same policy, voters will choose to discipline corrupt politicians and vote for an alternative party.

According to the instrumentalist theory, voters may also be ready to compromise politician quality for possible rewards. These could be vote-buying, patronage, or expected performance by corrupt candidates (Kramon, 2013). Some theoretical models and experimental and observational research support these predictions (De Vries & Solaz, 2017; Kramon, 2013). Nonetheless, the evidence is inconclusive as some experimental studies have also found no evidence of such trade-offs (Winters & Weitz-Shapiro, 2013). Other scholars have argued that voters need credible information to assign responsibility for corrupt politicians to be electorally punished (Besley, 2005). The evidence is again mixed. Experimental studies in Spain show that priming group identities reduce electoral punishment of corrupt politicians even when voters are informed about corruption and alternatives exist. Voters tolerate corrupt politicians because they offer higher expected utility than out-group candidates (Solaz et al., 2018). Other experimental evidence from real elections in Brazil shows that audits exposing municipal corruption reduce the chances of corrupt incumbent local governments being re-elected (Ferraz & Finan, 2011).

Based on this discussion, we expect that informing voters their coethnic MP was corrupt will not elicit any electoral penalty, given the benefits voters obtain from their identity group. In contrast, given performance expectations and alternative parties in legislative elections, we expect voters will be more likely to punish corrupt MPs when prompted to view them as partisans.

Hypothesis 3

After learning that politicians misused public goods, suggesting corruption, voters will be less likely to re-elect partisan MPs, whether copartisan or non-copartisan, than coethnics.

Setting

As in other African countries, the allocation of public funds in post-independent Kenya remains driven by patron-client networks. The incumbent President and central government have rewarded coethnics, hence respective areas of origin, with more public goods such as health, education services (Franck & Rainer, 2012), and roads, particularly in non-democratic periods (Burgess et al., 2015). Over the last two decades, Kenya has reformed its formal institutions, including adopting a new constitution in 2010, devolving power away from the presidency to make the system less winner-takes-all (Cheeseman et al., 2019). Still, informal institutions remain despite these reforms. Politicians continue mobilising support through ethnic patronage politics, portraying themselves as the best representatives of their constituents’ interests (Cheeseman et al., 2019).

Kenya has also renewed interest in decentralising funds to tackle the regional imbalances brought by patronage politics. The Constituency Development Fund (CDF) is the prime example of these new decentralisation efforts. The CDF channels central government funds to constituencies for the local MPs in collaboration with residents to decide which development projects to implement to reduce poverty at the grassroots level. Since 2003, the CDF has been Kenya’s principal poverty-reduction programme. The CDF receives no less than 2.5% of the government’s annual regular revenue by law, which translates into two million US dollars annually per constituency. The CDF finances five types of public goods: education, health, water supply, roads and bridges, and electrification. In practice, the fund has focused on education. During 2008–2013, the CDF spent about 50% of its expenditure on education projects like improving infrastructure, creating schools and classrooms, and providing primary, intermediate, tertiary, and postgraduate bursaries. These bursaries cannot exceed 5% of each constituency’s annual budget.

To spend allocations, the elected MPs must appoint and convene a local CDF committee composed of at most 15 non-paid residents. In collaboration with residents, these committees establish a priority list of projects to be implemented in their constituency. The committees send their top-priority projects to a district committee which reviews whether the projects align with the regulations. If so, a national committee distributes funds for the projects to be implemented with residents’ participation.Footnote 3

During the first administration of the CDF, 2003–2007, there were concerns about poor governance. MPs acted as the fund’s legislators, implementers, and auditors, yet several neglected to record its spending. Some MPs also used the CDF for what it seems are electoral purposes (Ngigi, 2015). For instance, during the election year of 2007, the number of CDF projects surged by 119%, which cannot be explained by funding increases of 38% in that year (Gutiérrez-Romero, 2013). The extra spending ahead of elections suggests that voters care about performance and that politicians have incentives to perform well as it can affect the probability of being elected. Since then, an independent government authority monitors the CDF, and MPs can be penalised for misuse. Still, in some constituencies, CDF projects have stalled, are of poor quality, or have fnot been completed as reported (World Bank, 2013).

Early studies analysing how the CDF was first implemented in 2003–2007 also concluded that only a minority of residents, 10–20%, had been consulted about how the CDF should be spent (Mapesa & Kibua, 2006). This figure has since then improved. According to the survey conducted for this article, nearly 40% of residents have participated in discussions about how to spend the CDF. This relatively high citizen involvement in how the CDF is spent makes this programme ideal for testing our hypotheses.

Data

We conducted a nationally representative survey in Kenya to understand perceptions about the Constituency Development Fund (CDF). The survey, carried out in December 2013, interviewed 1,210 people of voting age in 80 out of 290 constituencies across the country.Footnote 4 The sample was stratified by province, urban, and rural levels. All respondents were interviewed face-to-face, and their main characteristics are shown in Table A.1 in the Online Appendix.

Spending Preferences and Perceptions About the CDF

The survey asked all respondents how they wished the CDF to have been spent in their constituency during the 2008–2013 administration and how they believed the CDF was actually spent. To measure these preferences and perceptions, respondents were asked:

‘MPs are given money through the CDF fund to reduce poverty in their constituency. Think that all the money your MP is given through the CDF is represented by ten beans. How many of these ten beans would you prefer your MP spends on the following: Education, health projects, water projects, roads and bridge projects, electrification, administration of the CDF fund, or any other project you may wish.’

After that, respondents were asked:

‘How do you think your MP spent the CDF over the past administration 2008–2013? Thinking that all the money your MP was given through the CDF in your constituency is represented by ten beans. How many of these ten beans do you think your former MPs spent on the following: Education, health projects, water projects, roads and bridge projects, electrification, administration of the CDF fund, any other project, for projects consumed for the personal gain of the MP as a form of corruption.’

Experiment

Respondents Learn How Their MP Spent the CDF

The survey continued by asking respondents whether they wished to learn about how the CDF was spent in their constituency. Nearly 90% of respondents agreed to hear more information about the CDF. The characteristics of these respondents are, on average, the same as the overall sample, presenting no bias, as shown in Table A.1 in the Online Appendix.

Then we proceeded with our survey experiment by informing respondents how their previous MP’s administration had spent the CDF during the 2008–2013 period, according to official records. We shared the official CDF expenditure records whilst also randomising information about MPs. We reminded voters either that (1) the politician was an MP, (2) the name of the MP, hence hinting her/his ethnicity or (3) the political party of the MP. In Kenya, it is common that only a minority of MPs choose to contend for re-election. From our 80 sampled constituencies, 31 MPs contended for re-election as MPs, and only 19 of them won.Footnote 5 Only for these 19 constituencies respondents were informed that their MP had recently been re-elected. Respondents in all the other constituencies were told that the information provided about the CDF was about their former MP.Footnote 6

We randomised these attributes by asking each respondent:

‘Now, I’m going to tell you a little bit more about how the CDF was spent in your constituency during 2008–2013. But first, I would like you to pick any number between one and three.’

Immediately after, the interviewer provided the respondent with a showcard summarising how the MP had officially spent the CDF during the administration that had just ended during 2008–2013 in the respondent’s constituency, as shown in Fig. 1. The interviewer also read the following script aloud, depending on which number the respondent had randomly chosen.

Script 1: ‘Your MP [the statement ‘who was recently re-elected as MP’ was also mentioned in constituencies where the MP had been re-elected] spent the CDF in your constituency over 2008–2013, in the following percentages [interviewer read out load the percentages spent in each category] in education, health, water, roads and bridges, electrification, administration of the fund.’

Script 2: Same as script 1, and respondents were reminded of the name of the MP.

Script 3: Same as script 1, and respondents were reminded of the political party of the MP.

Fig. 1
figure 1

Type of card shown to survey respondents about how their MP had spent the CDF over 2008–2013

Full size image

In Script 2, we mentioned the name of the MP instead of their ethnicity to reduce a potential social desirability bias. Since most surnames in Kenya are linked to specific ethnic groups, reminding respondents about the name of their MP will provide a strong cue for the MP’s ethnicity. Immediately after the respondents were informed about how their MP had spent the CDF, we asked them:

‘Based on this information, do you agree or disagree with how your MP spent the CDF in your constituency?

Now that you know how the MP of this area spent the CDF over the last administration, how likely would you vote for her/him to become MP of your constituency again if you had the chance? More or less likely?’

Audit Results

The reported CDF expenditure might not reveal much about the quality of projects implemented or potential corruption. For this reason, the experiment continued by sharing with respondents the findings of independent audits about the CDF. A local NGO sponsored by UK-Aid independently conducted these audits. From the 80 constituencies sampled in our analysis, only 22 had been audited. Only for these audited constituencies were respondents read the additional script:

‘The National Taxpayers Association in Kenya, an independent and non-partisan organisation, found that in your constituency [x] % of the total CDF funds was wasted on badly implemented projects in which the fund was investigated in your area.’

To conclude the experiment, respondents were then asked:

‘Based on this new information how likely is that you would vote for him/her to become MP of your constituency again? More or less likely?’

Results

Preference, Perceptions and CDF Spending

We start by analysing respondents’ responses about how they wish the CDF to have been spent, by type of public good, and their perceptions about how the CDF was spent in their constituency during 2008–2013.

Education is the area where most respondents stated they would prefer the CDF to be spent on the most in their constituency, followed by health, water, and roads and bridges (Fig. 2, left-hand column). Respondents also believe that their MP spent the most on education projects (Fig. 2, middle column). However, the perceived expenditure on education is lower than the respondents’ preferred level of spending (23% vs 30% of CDF expenditure). Respondents also believe that about 12% of the CDF was lost in corruption for MP’s private gain. This figure is similar to the findings of independent and random audits, which suggest that 10% of the CDF was lost due to misuse.

Fig. 2
figure 2

Respondents’ preferences, perceptions and official records of CDF spending

Full size image

At this stage in the survey, we had not yet shown respondents how much officially their MPs had spent by type of CDF projects during 2008–2013. However, for comparison purposes, we have added that information in Fig. 2, in the right-hand column. According to these records, MPs spent the most on education projects. But this expenditure was, in fact, much higher (53% of the fund) than the respondents’ stated preference and perceptions about how the fund was spent.

Figure A.2, in the Online Appendix, disaggregates by ethnic groups the preferences about how the CDF should be spent and perceptions about how it was spent on each major type of project. This finding is important as preferences for in-group candidates, for instance, cannot be justified alone by different ethnic groups having different preferences on spending of the CDF. Most ethnic groups prefer to spend the most money on education projects, followed by health and then water projects. Similarly, each ethnic group believes their MP spent less on education than their preferred level. However, if we compare with the official records on CDF spending, MPs spent more than each ethnic group’s preferred and perceived spending level on education. That is, how the CDF was officially spent does not match respondents’ preferences or perceptions about how the fund was spent.

Are Coethnics and Copartisans Spending More in Line with Respondents’ Preferences?

We move on to test Hypothesis 1, whether, in the absence of any other information, voters believe that politicians with whom they have an affective tie (ethnicity or partisan affiliation) spend the closest to their preferred level of public goods than other politicians. We run a series of Ordinary Least Squares (OLS) regressions to test this hypothesis, as shown in Eq. (1).

$$difference_{ijk} = \alpha_{0} + \alpha_{1} E_{ij} + \alpha_{2 } X_{ij} + v_{ijk}$$
(1)

where difference represents a series of dependent variables, which measure the difference between the percentage that the respondent i residing in constituency j stated wished their MPs to have spent on the project k and the respondent’s perception about how the CDF was spent in her/his constituency. k denotes spending of the CDF on projects related to education, health, water, roads, electrification, administration of the fund, or others. Each of these dependent variables is continuous and ranges from 0 to 1. E is a dummy variable indicating whether the respondents living in constituency j where their MP had been recently re-elected in the 2013 election. We add this dummy in case candidates that got re-elected have smaller differences between respondents’ preferred and perceived spending levels. Vector X represents respondent characteristics, including respondents’ sex, whether they reside in a rural or urban area, age, number of children, whether they are employed, whether they have secondary or higher education attainment, an index of household wealth based on a list of 15 durable assets, and whether the respondent had indicated that they had gone without food to eat in the previous year for lack of funds.Footnote 7 We also added a dummy variable indicating whether the respondent belongs to the same ethnic group as their MP. We also added another dummy variable indicating whether the respondent is a copartisan of their MP. It is important to note that at this stage in the survey, respondents have not been reminded that their MP is a coethnic or copartisan. We simply constructed these variables by comparing the ethnic origin of the respondent and their MP and whether the respondent had voted in the previous election of 2007 for the same party to which her/his MP was affiliated to.

Table 1 shows that the difference between respondents’ preferred level of public spending and the perception of how the CDF was spent is not affected by whether the voters are a coethnic or copartisan of their MP. In other words, the gap in respondents’ preference and perception of how the CDF was spent is not driven by whether respondents’ MP shares an affective tie (coethnic or copartisan affiliation) with the respondents. That is the case across all the types of projects supported by the CDF.

Table 1 Difference between respondent’s preference and perception about how the MP spent the CDF by type of project
Full size table

As seen in Table 1, other factors explain the gap between respondents’ preferences about how the CDF should be spent and their perception of how the CDF was spent. For instance, respondents with a higher asset index and those living in urban areas and unemployed have a wider difference between their preference and perception about how much their MP spent on health services. Of course, this evidence does not rule out that MPs might have spent more on some projects and more in line with the desired level of spending of their coethnic and copartisan voters. After all, we showed earlier in Fig. 2 that voters largely underestimate how much their MP had spent on education CDF if compared to official records on spending. Thus, in Table 2, columns 1–7, we re-run our specification but using as a dependent variable the difference between the official CDF spending record and respondents’ preference for public spending for each type of project supported by the CDF.

Table 2 Difference between official CDF spending and respondent’s perception about how the MP spent by type of project
Full size table

Table 2, column 3, shows respondents whose MP is a coethnic have a wider difference between CDF spending and respondents’ preference for spending on water projects than respondents whose MP is a non-coethnic. Nonetheless, Table 2 reveals that there is no other spending bias in other types of projects if compared to the respondent’s preferences. Thus, we do not find evidence to support Hypothesis 1.

Figure A.3 in the Online Appendix shows there are no major differences in that explain differences across voter preference about how to spend the CDF.Footnote 8 In this analysis, we use the same specification as Eq. (1), using as the dependent variable the percentage that the respondent i residing in constituency j stated wished their MPs to have spent on the project k.

Thus far, the findings suggest that voters have very similar views about how they wish to spend the CDF across each of the main projects. Moreover, whether the MP is a coethnic or copartisan does not impact the differences between official expenditure and voter spending preferences or perceptions of how the MP spent the fund.

Results of Experiment

Having asked respondents about their preferences and perceptions about CDF spending, we then shared with them the official records on CDF spending in their constituency to assess how MPs’ attributes impacted their assessment of this spending. We did so by randomising MPs’ attributes. Respondents were reminded either that (1) the politician was an MP, (2) the name of the MP, hence hinting her/his ethnicity or (3) the political party of the MP.

Based on these three scripts and the respondents’ characteristics, respondents are divided further into five treatment groups. These treatments are those who: only heard the political role of their MP; those who heard that their MP is a non-coethnic; or a coethnic; and those that heard that the MP is a non-copartisan; and a copartisan. Table 3 shows the treatments are balanced among the three scripts we randomly assigned (informing whether the politician was an MP, her/his name, and political affiliation). Figures A.4 and A.5, in the Online Appendix, show that the respondents’ characteristics and how the MPs spent the CDF are well balanced across all our treatments.

Table 3 Treatments among survey participants that agreed to hear further information about how their MP spent the CDF in their constituency during 2008–2013
Full size table

We use an OLS regression, as shown in Eq. (2), to test the effect of the MPs’ attributes. The outcome of interest is whether voters disagreed with how their MP spent the CDF. This binary dependent variable is regressed on each level of the candidates’ attributes, omitting the reference categories, which serve as our ‘treatments’.

$$disagree_{ij} = \beta_{0} + \beta_{1} C_{ij} + \beta_{2} E_{ij} + \beta_{3} C_{ij} * \, E_{ij} + \beta_{4} D_{ij} + \beta_{5} X_{ij} + \varepsilon_{ij}$$
(2)

We code our dependent variable, disagree, as 1 if the respondent i residing in constituency j stated to have disagreed or strongly disagreed in how her/his MP spent the CDF fund, and 0 otherwise. The categorical variable, C, refers to the randomised attributes of the MP. This categorical variable has five levels. The first level, which serves as the reference category, denotes whether the politician is recalled just as an MP. The second and third levels are whether the name of the MP is mentioned and is a non-coethnic of the respondent or a coethnic of the respondent. The fourth and fifth levels are whether the name of the political party of the MP is mentioned and the respondent is a non-copartisan or copartisan of that party. E is a dummy variable indicating whether the respondents living in constituency j were told that their MP had been recently re-elected in the 2013 election. We include this dummy to see if re-elected MPs receive a premium in how voters evaluate their performance compared to those who are no longer sitting MPs. We also add the full set of interaction terms between these two categorical variables, C and E.

Constituencies receive and spend, on average, similar CDF funds. Thus to distinguish whether politicians are distributing the fund in alignment with voters’ preferences, we add vector Dij. This vector measures the difference in percentage points between the official CDF spending and the stated preference about how the respondent wishes the CDF to be spent on each designated CDF project (education, health, education, etc.). We include this information because it reflects voters’ calculations when evaluating politicians, according to the instrumentalist theory. According to this theory, voters prefer politicians who distribute the spending on public goods in close alignment with their preferred funding allocation.Footnote 9

We estimate two specifications of Eq. (2) by adding or not adding vector X, representing respondent characteristics. We use the same characteristics as in the previous regression specification. However, instead of adding two separate variables identifying coethnic and copartisan support, we add a single dummy variable indicating whether the respondent belongs to an ethnic group that supported Raila Odinga or Uhuru Kenyatta in the 2013 presidential election.Footnote 10 This variable helps us to distinguish more clearly how the electorate has been divided in elections, those supporting the opposition ODM party, or the incumbent government. We include X to decrease any potential biases introduced by chance and to estimate the standard errors more efficiently. The standard errors εij are clustered at the constituency level.

We graphically show the effects of the MPs’ attributes and the interaction between MPs’ attributes in Fig. 3. When comparing each attribute to the reference category, these effects represent the difference in the likelihood of a respondent disagreeing with how their MP spent the CDF fund. Respondents who were reminded that the politician was an MP fall into this reference category. Figure 3 shows the point estimates as dots, the 95 percent confidence intervals as lines, and the reference categories as dots without lines.

Fig. 3
figure 3

Probability of respondents disagreeing with how their MP spent the CDF after learning about reported CDF expenditure

Full size image

Figure 3, panel A, shows that priming respondents to think of their MPs as partisans increases their disapproval about how the MP spent the fund. For example, respondents who were reminded of their MP’s political party and are non-copartisan of the politician increased their disagreement with how the MP spent the fund by nearly 14 percentage points when compared to an MP who has no other information other than her/his political position. This effect is statistically significant. Similarly, respondents who were reminded that their MP is a copartisan increased their disagreement about how the MP spent the fund by nearly 20 percentage points, compared to an MP for whom there is no other information other than her/his political position.Footnote 11 A different story emerges for respondents who were reminded of the names of their MPs, thereby revealing their ethnicity. Regardless of whether the respondents are coethnic or non-coethnics, these respondents do not have any increased disagreement about how the MP spent the fund than the reference category.

Figure 3, panel A, also shows that the gap between the reported CDF expenditure in each of the main types of projects and how the respondents preferred to spend the fund has no effect on the likelihood of disagreeing with how the MP spent the fund. It makes no difference either if the politician was recently re-elected or not.

In sum, regardless of how their MP officially spent the fund versus respondents’ preferred spending level, respondents disagree with their MP’s spending when reminded that their MP is a partisan (whether copartisan or non-copartisan) but not when reminded that their MP is a coethnic. These results remain robust when the characteristics of voters, vector X, are included, as shown in Fig. 3 panel B.

One potential concern with our current analysis is whether respondents understood the information we provided about how their MP spent the fund. For that reason, we also asked respondents why they had agreed or disagreed with the spending made by their MP. Reassuringly, as shown in Fig. A.6, a significant percentage (24%) said that their MP had spent (way) too much on education, while another 10% did not believe the MP had spent that amount.

Next we directly test Hypotheses 2a and 2b, whether respondents would still re-elect their coethnic and partisan MPs once learning about official CDF spending in their constituency. We use the same regression specification shown earlier. However, the dependent variable now takes the value of 1 if respondents said they agreed or strongly agreed that they would vote again for the same MP, and 0 otherwise.

Figure 4 shows that respondents whose MP is a coethnic increase the probability of claiming that they would still vote for the same politician by nearly 14 percentage points, compared to an MP for whom there is no information other than her/his political position. Interestingly, no other MP’s attributes influence the probability of voting in the future for the same MP compared to the reference group (Fig. 4).

Fig. 4
figure 4

The probability that respondents claim that they would vote again for their MP after learning about reported CDF expenditure

Full size image

Even though respondents reject that ethnicity is an attribute they consider important when considering who to vote for if asked openly (Fig. A.1), our experiment suggests that coethnic MPs have an advantage in terms of re-election. Moreover, the probability of re-electing or not an MP is not affected by the disparity between how the CDF was spent and how the respondent preferred the CDF was spent. These effects remain robust if adding voter characteristics that have been argued could explain ethnic voting behaviour, such as educational attainment and wealth level (Kramon, 2013), as shown in Fig. 4 panel B. Thus, we find support for both Hypotheses 2a and 2b.

Audits

The survey experiment ended by showing the respondents evidence of CDF misuse based on independent audits. Only 22 of the 80 constituencies in our survey were audited. These audits found that 10% of CDF spending was wasted on poorly implemented projects, with a couple of constituencies having zero or 40 percent of poorly executed projects. To test Hypothesis 3, whether voters sanction politicians for mismanagement and potential corruption, we use the regression specification shown in Eq. (3).

$$voteagain_{ij} = \rho_{0} + \rho_{1} C_{ij} + \rho_{2} A_{ij} + \rho_{3} C_{ij} * \, A_{ij} + \rho_{4} E_{ij} + \rho_{5} C_{ij} * \, E_{ij} + \rho_{6} D_{ij} + \rho_{7} X_{ij} + u_{ij}$$
(3)

The dependent variable voteagain takes the value of 1 if respondent i residing in constituency j stated that they will vote again for the same MP after hearing the results of the independent audits of the CDF, and 0 otherwise. The continuous variable A represents the percentage of misused found in the audits. The impact of the audit information, our new treatment, on respondents’ voting choice is captured by the interaction between the MP’s attributes and the percentage of funds misused, Cij* Aij. All the rest of the vectors are the same as in the previous specification. As before, we control for whether the MP was re-elected or not because other studies have found that politicians that have re-election incentives misuse fewer resources than those without re-election incentives (Ferraz & Finan, 2011). The standard errors uij are clustered at the constituency level.

As previously stated, the spending of the CDF is balanced across our treatments (Fig. A.5). However, there are significant imbalances in the sub-sample of 22 constituencies that were audited. We change the reference group in our analysis to make it easier to see these biases across our treatments and directly test Hypothesis 3. This time, as shown in Fig. 5, we use respondents who were reminded that their MP is a coethnic as a reference group. Compared to that reference group, the audits showed a lower percentage of misused funds for respondents who were only mentioned the political role of their MP, whose MP is a non-coethnic MP, and non-copartisan MP (five, nine and ten percentage points difference). As Fig. 5 also shows, coethnic and copartisan MPs misused funds roughly at the same rate.

Fig. 5
figure 5

Percentage of misuse of CDF found by independent and random audits by treatment

Full size image

Table 4 shows that the more funds partisan MPs misused, whether copartisan or non-copartisan, the less likely respondents are to re-elect them compared to a coethnic MP. These effects are statistically significant. In contrast, non-coethnic MPs suffer no electoral sanction compared to coethnic candidates.Footnote 12 These contrasting electoral sanctions cannot be explained solely by differences in audit findings because non-coethnics and non-copartisans have a similar percentage of funds missing, which is lower than that of coethnic MPs. Thus, simply informing respondents of the MP’s partisan affiliation (whether copartisan or not) undermines the assessment of MPs in a way that informing respondents of ethnicity does not, supporting Hypothesis 3.

Table 4 Probability that respondents would vote for their MP after learning the results of independent audits on misuse of the CDF in their constituency
Full size table

The interaction Cij* Eij suggests MPs who were re-elected in 2013 and misused CDF funds have no advantage in terms of respondents saying they would vote for them again. Given the smaller sample size of audited constituencies (270 respondents), this result may be due to a lack of statistical power, but it is consistent with our previous findings that recently re-elected MPs have no electoral advantage in our experiment.

Conclusion

According to the instrumental theory, even if ethnicity or partisanship are used to determine which candidate is expected to provide better access to public goods if elected, voters will update their expectations as more reliable information becomes available, such as on politicians’ performance (Carlson, 2016; Conroy-Krutz, 2013). We tested these assumptions with a survey experiment in Kenya. Given the complexities of government spending, we concentrated on voters’ perceptions of the CDF, the country’s flagship poverty-reduction programme, in which each MP receives roughly the same funds and decides which types of public goods to finance after consulting citizens. The CDF spending is widely publicised, thus ideal for testing voting theories’ assumptions. Similar programmes exist in Africa, Asia, and Latin America. Our findings could therefore help explain voting behaviour in these similar settings.

According to our findings, voters have incorrect perceptions of how their incumbent MP spent public funds in their community. These misconceptions are unaffected by MPs’ ethnicity or partisanship. A second significant finding is that partisanship (whether copartisan or not) influences how voters evaluate politicians’ performance in ways that ethnicity does not. Voters’ approval and likelihood to vote for their MPs decreased after they were informed about how their MPs spent public funds, including audit results, and prompted to consider their MPs as partisans. Despite poor performance and evidence of corruption, coethnic MPs were more likely to be re-elected. These contrasting evaluations emerged even though the audits revealed wide-ranging misappropriation of finances by MPs, and non-copartisans mismanaging fewer sums than coethnic MPs.

Kenya has introduced several reforms to devolve power away from the president and strengthen political competition. Fierce competition in Kenyan legislative elections, as in other similar countries, may explain why partisanship is linked to performance and voters are willing to punish poor governance if primed to think about their candidates in partisan terms (Barkan, 2009).

Despite these formal institutional reforms, Kenyan informal norms enable politicians to mobilise supporters along ethnic lines (Cheeseman et al., 2019). These mobilisation tactics may lead voters to perceive their coethnic candidates as being the best to represent their interests, prioritising group loyalties over bad or corrupt performance, as the identification and expressive voting theories suggest. One could also argue that our findings might be driven by decades of patron-client networks where incumbents’ coethnics enjoyed patronage benefits at the expense of others (Burgess et al., 2015). Although favouritism may take many forms, we found no evidence that MPs’ spending decisions match the spending priorities of coethnics or copartisans. There is no evidence either to suggest ethnic groups have different public spending preferences that could justify voters preferring coethnic candidates. Overall, these findings contribute to discussions about the relevance of voting theories in nascent democracies.