---
title: "Is the IMF a Scapegoat? A Survey Experiment in Kenya & Ghana"
authors: ["Matthew DiGiuseppe", "Kathleen J. Brown", "M. Rodwan Abouharb", "Bernhard Reinsberg"]
year: 2024
status: "Working paper (recently accepted)"
venue: "Conditionally accepted at International Studies Quarterly"
url: "https://www.matthewdigiuseppe.com/papers/imf-scapegoat-survey-experiment-kenya.md"
links: {"PDF on PEIO": "https://www.peio.me/wp-content/uploads/PEIO16/submission_76.pdf"}
full_text: true
---

# Is the IMF a Scapegoat? A Survey Experiment in Kenya & Ghana

DiGiuseppe, M., Brown, K., Abouharb, R., & Resinsberg, B. (2024). Is the IMF a Scapegoat? A Survey Experiment in Kenya & Ghana. Conditionally accepted at International Studies Quarterly.

- Status: Working paper (recently accepted)
- PDF on PEIO: https://www.peio.me/wp-content/uploads/PEIO16/submission_76.pdf
- Listed on: https://www.matthewdigiuseppe.com/#research

<!-- END OF GENERATED HEADER: edit freely below this line; scripts/build_agent_files.py keeps it -->

## Abstract

Many scholars and practitioners have proposed that leaders ‘scapegoat’ the IMF for unpalatable economic reform programs often undertaken during times of economic crisis. However, there is limited evidence that these tactics are effective in shifting citizens’ blame onto the IMF, or that they protect the government from public protest or electoral punishment. This research note seeks to test the micro-foundations of the scapegoat hypothesis with a high-powered online survey experiment in Kenya. Our study provides two insights. First, we find that blaming the IMF has small but significant effects on incumbent support. However, this effect is insignificant when respondents are also exposed to counter-narratives in which the opposition blames the government. Second, we find that blame is not zero-sum. While attributing blame to the IMF increases respondents’ own blame attribution to the IMF, it does not significantly reduce the blame placed on the incumbent. Overall, the results question the political utility of the scapegoating strategy.

## Full text

> Extracted automatically from the conference version dated June 4, 2024: https://www.peio.me/wp-content/uploads/PEIO16/submission_76.pdf. Tables, figures and equations may be garbled or missing; quote the PDF, not this text.

**Is the IMF a Scapegoat? A survey experiment in Kenya**

Rodwan Abouharb · Kathleen J. Brown University College London · Leiden University Matthew DiGiuseppe · Bernhard Reinsberg Leiden University* · University of Glasgow June 4, 2024

Abstract Many scholars and practitioners have proposed that leaders ‘scapegoat’ the IMF for unpalatable economic reform programs often undertaken during times of economic crisis. However, there is limited evidence that these tactics are effective in shifting citizens’ blame onto the IMF, or that they protect the government from public protest or electoral punishment. This research note seeks to test the micro-foundations of the scapegoat hypothesis with a high-powered online survey experiment in Kenya. Our study provides two insights. First, we find that blaming the IMF has small but significant effects on incumbent support. However, this effect is insignificant when respondents are also exposed to counter-narratives in which the opposition blames the government. Second, we find that blame is not zero-sum. While attributing blame to the IMF increases respondents’ own blame attribution to the IMF, it does not significantly reduce the blame placed on the incumbent. Overall, the results question the political utility of the scapegoating strategy.

*Corresponding author: [email removed] This research is funded by European Research Council H2020 Grant #852334.

Governments borrowing from the International Monetary Fund (IMF) frequently resort to blaming the Fund for implementing unpopular economic reforms. A senior IMF official noted that blaming the IMF is a recurring pattern in numerous countries, often serving as a politically convenient narrative. This official emphasized the consequential nature of scapegoating the IMF had on the public’s view of the organisation, making its work challenging and requiring resilience (Interview with senior IMF official, 2022). This anecdotal evidence reflects a long-standing hypothesis in the literature on IMF lending, first proposed by Remmer (1986), that leaders blame the Fund for costly economic reforms, thereby avoiding public backlash that may endanger their tenure in office. A government may prefer to consolidate public debt and reforms may be imminently necessary to avoid a crisis, but strong opposition from the public and interest groups can threaten the government’s survival should it implement harsh tax increases and spending cuts. In this case, a government may invite the intervention of the IMF and attempt to shift the blame for reforms onto the Fund. By framing the Fund as the driver of reforms and promoting a narrative that it is powerless to refuse IMF intervention, the government may achieve fiscal stabilization without paying the price of public support.

Since Remmer (1986), the scapegoating hypothesis has become a well-established argument in IMF research (Vaubel, 1986; Edwards and Santaella, 1993; Vreeland, 2003; Bird and Willett, 2004). Some research has argued that governments seeking to avoid blame for unpalatable economic policy changes is a key reason they agree to an IMF reform program (Dreher and Walter, 2010; Rickard and Caraway, 2019). Empirical analysis of the efficacy of scapegoating has used national-level proxies such as election outcomes or sought to extract insights from observational public opinion surveys (Dreher and Vaubel, 2004; Smith and Vreeland, 2006; Alca˜niz and Hellwig, 2011; Williams, 2012). While these existing studies are valuable in helping us to understand the underlying logic of government’s scapegoat strategies, this literature has yet to directly assess the efficacy of these strategies on the voting intentions of the public. Studies that look at leader survival assume that scapegoating is a possible mechanism (Smith and Vreeland, 2006; Williams, 2012) without ruling out other mechanisms. Studies that show citizens blame international actors, like the IMF, for economic conditions in observational public opinion surveys (Alca˜niz and Hellwig, 2011) do not demonstrate that this effect is due to government efforts to blame the IMF or that it is causally linked to incumbent support.[^1]

While informative, these studies tell us little about the susceptibility of voters to IMF scapegoating narratives and if these blame-shifting strategies can influence individual-level outcomes like support for the incumbent or voting intention in their next election. As such, our understanding of how effective governments are in displacing responsibility to the IMF remains limited. Yet, understanding IMF scapegoating is crucial, as the legitimacy of reform programs hinges on the perceptions of those affected by them, and the political processes through which they are implemented (Ortiz and B´ejar, 2013; Konstantinidis and Reinsberg, 2023; Bratton et al., 2005). Additionally, the governments’ ability to effectively shift responsibility for these policies to the IMF has substantial implications for their democratic accountability to their citizens. Lastly, existing research of IMF reform program lending are often built on the assumption that scapegoating is an important part in political decisionmaking (Dreher and Walter, 2010; Vreeland, 1999; Dreher and Gassebner, 2012; Kern et al., 2023). Finding evidence for the core assumptions of these models can increase confidence in their applicability.

To get closer to understanding the effectiveness of scapegoating, we turn to a survey experiment. This approach allows us to assesses the effectiveness of scapegoating at the individual level and overcome the limitations of observational data in establishing causality. By having greater control of the information presented to respondents, it also allows us to consider multiple blame narratives that emerge in a competitive political environment. Specifically, it also allows us to consider the possibility that opposition parties may attempt to counter government narratives and reduce their effectiveness. Lastly, an experiment permits the measurement of links in the causal chain from IMF blame narratives to incumbent support and reelection that is not possible when researchers can only examine leader survival or government approval data from pre-existing surveys.

Below, we present the findings from an original online survey experiment fielded on 2,100 Kenyan citizens in December 2023 and January 2024, representative of the population in terms of age, gender, and region, recruited via a Facebook advertising campaign. The analysis reveals two key insights that questions the utility of blaming the IMF for unpopular economic reforms. First, we find that a fictional news article blaming the IMF for economic reforms has a small but significant effect—10% of a standard deviation—on support for the Kenyan incumbent William Ruto. However, when respondents are also presented with a counter-narrative in which the opposition blames the incumbent, the treatment effect size drops in half and is no longer significant. This finding raises questions about the ability of government to blame the IMF in the presence of an opposition that will likely want to blame the incumbent.

Second, our exploratory analysis finds that blame is not necessarily zero-sum. Our main treatment was successful at significantly increasing blame for the IMF. Yet, we find that when asked about blame for the incumbent, the treatment did not move attitudes. The implicit assumption of existing scapegoating research (Dreher and Walter, 2010; Vreeland, 1999; Dreher and Gassebner, 2012; Kern et al., 2023) that increasing blame for the IMF reduces blame for the incumbent may not be accurate. There appears to be enough blame to go around in Kenya.

### Theoretical Framework

We draw upon the existing literature examining politicians scapegoating international organisations (IOs) and the literature on economic voting to provide a more developed set of expectations about the efficacy of shifting blame upon the IMF with respect to incumbent support. A wide-ranging literature has argued that member states utilise IOs to deflect blame for unpopular domestic policies (Abbott and Snidal, 1998; Dreher et al., 2018; Louis and Maertens, 2021; Moravcsik, 1994; Tallberg, 2002; Thatcher and Sweet, 2002; Vaubel, 1986; de Vries et al., 2021). The most theoretically developed literature has examined national politicians’ strategies (Weaver, 1986; Hood, 2011; Hinterleitner and Sager, 2017) frequently shifting blame onto the European Union (EU) and its institutions (Schlipphak and Treib, 2017; Sommer, 2020; Traber et al., 2020; Heinkelmann-Wild and Zangl, 2020). Politicians often take ‘presentational strategies’ by engaging in public communication that minimizes their own blame.

In terms of the effectiveness of these blame-shifting strategies, studies have found that governments can manipulate the views of European publics. These findings are supported by content analyses of press coverage of EU policy failures (Rittberger et al., 2017; Kriegmair et al., 2022), public opinion polls (Gabel and Scheve, 2007; Kalbhenn and Stracca, 2015; Schlipphak and Treib, 2017), and election outcomes (Hobolt et al., 2013; Lobo and Pannico, 2020; Biten et al., 2023). Further, experimental studies suggests that the effect of blame shifting is non-trivial (Kumlin, 2011; Maier et al., 2012; Hobolt et al., 2013). This research finds that factors such as the involvement of messengers in EU policies (Kriegmair et al., 2022), their trustworthiness in the eyes of the audience, as well as respondents’ education and skepticism of the EU (Hobolt et al., 2013) affect the extent to which messengers can manipulate public opinion.

As we mentioned above, the IMF reform literature indicates that governments may adopt similar strategies to scapegoat the IMF and avoid blame for costly structural adjustment (Remmer, 1986; Vaubel, 1986; Edwards and Santaella, 1993; Vreeland, 2003; Bird and Willett, 2004). We incorporate the insights from the literature on economic voting, which draws out how voters evaluate economic policy changes and apply it to the context of IMF reform programs. The broad consensus from the economic voting literature is that respondents overwhelmingly support economic policies that benefit them (Kaufman and Zuckermann, 1998; Bratton and Mattes, 2003; Hays et al., 2005; Walter, 2010; Margalit, 2013; Rickard, 2015; Jamal and Milner, 2019). IMF programs often entail painful economic reforms and could lower support for national politicians. Citizens may evaluate these programs through narrow pocketbook approaches assessing the costs and benefits these programs have on them and their families (Duch and Stevenson, 2006), or they may take a more expansive sociotropic view of economic reform programs’ broader costs or benefits to society (Kaufman and Zuckermann, 1998). To be sure, the existing empirical literature does not allow us to distinguish between citizens taking on an individual or broader societal view when assessing IMF programs. However, it is clear that IMF programs have sufficient economic costs that politicians will benefit from avoiding blame for their implementation.

Taken together, the broader blame-shifting literature and the economic voting literature suggest that governments blame-shifting the IMF for economic policy changes may be an effective strategy to increase popular support. This discussion leads to our hypotheses about the efficacy of blame-attribution on incumbent support in the context of IMF structural adjustment. The first hypothesis considers that blaming the IMF for economic reforms can have a successful scapegoating effect that protects the incumbent from negative voter reactions. This expectation is consistent with the blame-shifting literature. When responsibility for reforms is attributed to the IMF rather than to the incumbent, support for the incumbent should increase.

Hypothesis 1. Individuals will be more supportive of the incumbent when responsibility for structural adjustment is attributed to the IMF.

In reality, public discourse surrounding economic policy is unlikely to be so straightforward as to either blame government leadership or the IMF for the costs of structural adjustment. While the government may attempt to paint a narrative that the IMF is responsible, domestic actors, in competitive media and political environments, have the opportunity to counter government blame attribution to increase their electoral chances (Abouharb et al., 2023).

In a second hypothesis, we consider that the political opposition may attempt to undermine the incumbent’s scapegoat efforts and encourage voters to punish the government for costly economic policy. Here, we consider a scenario where the incumbent blames the IMF for reforms, but the opposition disputes IMF responsibility and instead blames the government for the reforms. In this divided narrative, support for the incumbent should be lower when the opposition provides a counter narrative and blames the government, compared to a situation when the government simply attributes responsibility to the IMF without opposition.

Hypothesis 2. Individuals will be less supportive of the incumbent when responsibility for structural adjustment is attributed to the IMF and when the opposition blames the government for reforms.

According to the IMF scapegoating hypothesis, the causal link between blame attribution and incumbent support is through reduced blame for the incumbent. For example, Dreher and Walter (2010, 4) suggest that because currency devaluations are highly unpopular, governments that want to implement them must invoke the Fund and use it as a scapegoat, thereby implementing their desired economic policies whilst ensuring their political survival. We demonstrate this path in Figure 1. We follow our central hypotheses by exploring if narratives that place blame on the IMF actually increase the public’s blame attribution directly on Fund, enabling incumbents to avoid blame. We expect that framing structural reforms as the responsibility of the IMF causes individuals to shift their blame attribution away from the incumbent for the specific implementation of the IMF policy, whilst approval for the incumbent increases.

Hypothesis 3. Individuals will be less likely to blame the incumbent for structural adjustment when responsibility is attributed to the IMF.

IMF blame attribution  Incumbent avoids blame

Incumbent approval

Figure 1: Causal path diagram from IMF blame attribution to Incumbent Approval

### Survey design

We test our hypotheses by fielding an online survey experiment in Kenya in late December 2023 and early January 2024. Kenya provides a useful setting for several reasons. First, Kenya was under an IMF program at the time of the survey, with conditions that required austerity measures including tax increases and spending cuts. Like most of the Global South at this time, the Kenyan population was wrestling with higher food and fuel prices due to global inflation and wars in Ukraine and the Middle East. As such, the austerity has proven politically contentious and sparked a number of violent protests by supporters of the opposition (Auvinen, 1996; Ortiz and B´ejar, 2013; Reinsberg et al., 2023).

Notably, the responsibility for the measures has largely been attributed to William Ruto, the incumbent, in the popular press. Further, Ruto has done little to blame the IMF for the measures when addressing the public or the media. For the purposes of testing the IMF scapegoating hypothesis, this serves two purposes. First, citizens are aware of the austerity and the impact it is having on their lives. As such, the survey touches on a topic all citizens have close familiarity with and gives them a motive to place blame. Second, there is little evidence that the public has been pre-treated with a narrative in which the IMF is to blame and Ruto is carrying out a policy that he does not support. Consequently, this provides an environment in which our experimental narrative has the opportunity to move attitudes.

To test our three hypotheses, respondents were randomly assigned to one of three treatment arms where they were shown fictional excerpts of newspaper articles. These articles are formatted to resemble news stories published online by The Daily Nation, Kenya’s largest English-language daily newspaper.[^2]

One-third of respondents assigned to the “control” group are shown an article, Figure 2, that describes the need for the Kenyan government to implement a new round of economic reforms. One-third are assigned to the “IMF blame” treatment and are shown a similar article, left panel in Figure 3, that describes the economic reforms as mandated by the IMF and that the government has little choice but to enact the reforms. One-third are assigned to the “IMF-government blame” treatment and are shown an article, right-panel Figure 3, that frames reforms as IMF-mandated, but also discusses how the political opposition attributes responsibility to the government.

Following the treatment articles, we asked our respondents whom they attribute blame for austerity measures. We asked about blame in a forced choice (IMF, Ruto, Don’t Know) but also asked about blame attribution independent of other targets: “How responsible do you think President William Ruto is for the recent tax increases and reduction in fuel subsidies?” (11-point scale, Not responsible - Completely responsible). The outcome of this question is

Figure 2: Control: The control is shown to 33% of respondents and preceded by this sentence: “This is an excerpt from a newspaper article published on December 20, 2023.

Please read it carefully. We will ask questions about the contents.”

used to test Hypothesis 3. We also asked a similar question about how much responsibility respondents attribute to the IMF as a further manipulation check. We expect those that receive the treatment to attribute more blame to the IMF than those in the control group.

Following the questions on blame attribution, we asked our respondents four questions about their assessment and vote intention for William Ruto. We use these four questions to create an additive index of Incumbent Support to test Hypotheses 1 and 2. These incumbent support questions are based on those constructed by Schleiter and Tavits (2018). The construction of an index aims to increase precision in estimation of individuals’ attitudes towards the incumbent by reducing measurement error. We add the responses to the questions below and standardize the scale so that it has a mean of 0 and standard deviation of (a)  (b)

Figure 3: Treatments: Each treatment article sent to 33% of respondents. They are preceded by this sentence: “This is an excerpt from a newspaper article published on December 20, 2023. Please read it carefully. We will ask questions about the contents.”

1. Higher values indicate higher approval.[^3]

We will now ask you for your opinion about President William Ruto. We want to know what you think! Please take your time and read each question carefully.

- “How would you rate the performance of President William Ruto?” 11-point scale from “Very negative” to “Very positive”

- “How would you rate the performance of President William Ruto in managing Kenya’s economy?” 11-point scale from “Very negative” to “Very positive”

- “Imagine that presidential elections will be held tomorrow. How likely would you be to vote for William Ruto?” 11-point scale from “Very unlikely” to “Very likely”

- “How well do you think President William Ruto will manage the Kenyan economy in the future?” 11-point scale from “Very badly” to “Very well”

S

ampling Following a power analysis (see Supplementary Appendix), we targeted a sample of 2,100 Kenyan adults recruited via a Facebook advertising campaign. The sample size is geared toward finding a minimal detectable effect of 15% of a standard deviation.[^4] Facebook users with accounts registered within Kenya were eligible to be shown an ad which invites them to participate in a short survey, in which they can earn mobile phone minutes (“airtime”) as compensation. The ad contained a link to the survey questionnaire, hosted in Qualtrics. Over 11,000 users clicked our link. After respondents were filtered out because of full quotas or other eligibility requirement and duplicate entries were deleted, we have a final sample of 1,946 respondents. We present the balance of demographics across treatment conditions and a comparison with census data in the Supplementary Appendix.

Social media recruitment has several advantages in the Kenyan context. A large share of the population has access to internet and has a social media account, with 98% of adults having access to a mobile phone (Kharono et al., 2022) and 49% with a Facebook account. While mobile phone and social media use skews towards wealthier, highly educated urban citizens, this is also a problem when recruiting online respondents from the panels of traditional survey companies in the region. Recruitment via Facebook allows access to a wide pool of respondents at a low cost. This sampling approach is similar to those previously used by Rosenzweig and Zhou (2021) and Pham et al. (2019) to recruit online survey respondents in sub-Saharan Africa. Further, providers of quota panels in the region use a similar recruitment process.

Our recruitment ad was shown to Facebook accounts registered in Kenya to users above 18 years of age and was displayed to users until 2,100 complete survey responses have been collected. Importantly, ads are not assigned randomly. Instead, Facebook shows ads to an audience most likely to click on the link within the ad as determined by the Facebook advertising auction system algorithm. Further, the ad was targeted separately to male and female users, and to users based on broad age categories derived from the results of the 2019 Kenyan census.[^5] This targeting approach follows the recommendations outlined in Neundorf and ¨Ozt¨urk (2023), where the authors find that maximizing link-clicks and targeting ads along broad demographic quotas like age and gender can improve representation in the sample.

To capture a sample reflective of the population, we set nationally representative quotas in Qualtrics for age, gender, and region. We screened out respondents after these quotas are filled. In line with previous survey research via social media in developing countries (Rosenzweig and Zhou, 2021; Pham et al., 2019), we expect our Facebook sample to be skewed towards urban, highly educated, and wealthier citizens and so will not attempt to achieve a nationally representative sample along these attributes. Further, the survey and recruitment was presented in English. English is a compulsory subject in Kenyan schools and an official language along with Swahili. Still, there maybe a population only reading Swahili or non-literate that is necessarily excluded from our quota sample. As such, the survey excludes the small amount of Kenyans that do not read English.

Respondents were compensated for their participation with Ksh. 100 (€0.60) airtime minutes via text message. While ethical, providing compensation creates incentives for multiple responses and sharing with like-minded friends. We used several strategies to ensure that individuals only take the survey once, and to guard against the survey link being redistributed outside of the Facebook advertising platform. First, we activated Qualtrics’ fraud detection, which attaches a cookie to the user’s browser, preventing the same device-browser combination from submitting subsequent survey responses. Second, we discarded responses that submit the same telephone number for compensation as has already been submitted in a previous response. Third, we discarded responses that had identical responses on four categories (IP Addresses, Sex, Region, Age). In developing countries it isn’t uncommon to have responses from the same IP addresses as mobile providers will use the same IP for multiple users. As such, we only discarded responses that were similar on additional dimensions.[^6] Lastly, we screened our respondents that reported that they received a survey link from anywhere other than a “Facebook advertisement”.

### Analysis

Figure 4 presents the average treatment effect (ATE) estimated by OLS and complier average causal effect (CACE) estimated using two-stage least squares. The later used the randomized treatment as an instrument for whether or not the respondent “recieved” the treatment. We codify this as being able to correctly identify the fictitious article name post-treatment. Estimating the CACE (sometimes referred to as a local average treatment effect) has the benefit of allowing us to test the effect on those that were paying attention without having to drop results based on post-treatment attention checks or attention checks unrelated to the treatment. However, its interpretation is different than the ATE. The CACE indicates the effect of the treatment only among those that ‘comply’ with or received the treatment while the ATE indicates the effect of the treatment among all respondents regardless of their reception of the information.

In both estimations, we included a matrix of pre-treatment covariates and their coefficients on the right hand side λD. The covariates help increase the precision of the estimate and thus the statistical power. We chose these variables agnostically with a LASSO selection model following the recommendation of Bloniarz et al. (2016). We first include each continuous variable and each categorical variable (as independent dummies) in a model predicting the outcome. The LASSO model returns only variables with non-zero coefficients. We then include these non-zero variables in our model estimating the treatment effects. This has the benefit of increasing the precision of our estimates (i.e. increasing power) without increasing researcher degrees of freedom and without overfitting the model with numerous covariates. Notably, we asked pre-treatment about respondents’ assessment of President Ruto (and other prominent politicians). This correlates strongly with our incumbent vote choice outcome (r[^2]=0.8). As such, it serves to substantially increase the power by allowing a “quasi pre-post” within subjects design (Clifford et al., 2021).

The top panel of figure 4 presents the results of two models estimating both the ATE and the CACE. In both cases the effect is small but significant. Since the dependent variable is standardized (mean=0, SD=1), both estimates can be interpreted as a percentage of a standard deviation. Compared to the control group, those shown an article blaming the IMF for reforms demonstrate 7% or 10% of a standard deviation higher support for the incumbent. These are small effects but it is important to keep in mind that the treatment is quite short and fleeting compared to repeated exposure to blame narrative citizens may experience in real life.

Hypothesis 1 − ATE/CACE on Ruto Support

ATE CACE

```text
−0.2  −0.1  0.0  0.1  0.2  0.3
```

Coefficient estimates and 95% confidence intervals

Hypothesis 2 − ATE/CACE on Ruto Support

ATE CACE

```text
−0.2  −0.1  0.0  0.1  0.2  0.3
```

Coefficient estimates and 95% confidence intervals

Figure 4: Tests of Hypothesis 1 and 2: Each panel presents the average treatment effect (ATE) and complier average causal effect (CACE) estimated with OLS and two-stage least squares respectively. The dots indicate the point estimates and the bars indicate the 95% confidence intervals. Each model is estimated with HC2 robust standard errors.

The bottom panel compares the control group to the group that received an article blaming the IMF and noting that the opposition blames the incumbent. Here we see that the already small point estimates are almost cut in half and the statistical significance is not retained. This suggests that the opposition blame can successfully counter the government’s narrative.[^7] This is consistent with framing research that shows opposing frames can neutralize each other (Chong and Druckman, 2010). As we show in the supplementary appendix, the effect in our study appears to hold for both previous voters of Ruto and the opposition alike. We found no substantive differences in the treatment effects across political support.

To further probe our results, we look at the ATE when estimating each component of the

Hypothesis 1 − ATE on Disaggregated Ruto Support

model

ATE−EconomyFuture ATE  ATE−Vote ATE−Economy ATE−Assess

```text
0.00  0.05  0.10  0.15
```

Coefficient estimates and 95% confidence intervals

Figure 5: ATE - Each Index Component: This figure presents the ATE from four models each predicting a different outcome of the incumbent approval index. The dots indicate the point estimates and the bars indicate the 95% confidence intervals. Each model is estimated with HC2 robust standard errors.

additive index of incumbent support independently. Figure 5 presents these findings. We see that the largest and most significant effect for the outcome positing a hypothetical vote for the incumbent, Ruto, in the next election. The ATEs for the remaining outcomes are insignificant, but in line with the direction and magnitude of the overall index. As such, we have grater confidence the the small effect sizes presented in Figure 4 are not due a negative relationship with one component of the index.[^8]

Next, we pay closer attention to the path from blame to support by looking directly at blame attribution for the economic reforms. Hypothesis 3 states expectations regarding blame of the incumbent as a function of the treatments. Recall that we asked two questions about blame. One asked about blame and gave a choice (IMF, Ruto, Don’t know). We simply coded this answer into a binary outcome ‘blame for IMF’ and ‘blame for Ruto’ leaving the remaining categories in the baseline. The panels on the right of Figure 6 present the estimates based on linear probability models comparing IMF blame treatment to the control condition. The second measure of blame results from two questions in which we inquired about blame for the reforms on a 0-10 scale for the IMF and Ruto separately. The results of these analyses can be seen on the left of Figure 6.

We see that when given a closed-choice, the treatment of blaming the IMF reduces blame placed on Ruto by 12%(ATE) - 20%(CACE). However, when asked to simply place blame for reforms on a 0-10 scale for Ruto and IMF independent of the other, the treatment does not have a statistically significant effect.[^9] Again, the scale is standardized, as such the effect sizes are interpreted as percent of a standard deviation. The null effect of blame on Ruto may be a function of the treatment not working (despite our results above). However, the bottom panels of 6 indicate that the treatment has a sizeable effect on blame towards the IMF in both the scale outcome (0.3-0.4) and the choice outcome (14 - 20%). Consequently, this is a strong indication that the treatment was successful and that the null effects on blame toward Ruto were not a function of a weak treatment.

We briefly draw attention to the comparison between the second treatment condition (IMF blame + counter narrative) and the control on the blame outcomes in Figure 7. The second treatment does not significantly change either ‘blame for Ruto’ or ‘blame for the IMF’ when using the scale outcomes. This is consistent with the test of the first hypothesis in which opposition blame attenuates the incumbent’s efforts to blame the IMF. We are reassured that the reduced treatment effect is due respondents reducing blame on the IMF. However, the treatment does impact blame when respondents are forced to choose between the IMF and Ruto, albeit with a smaller magnitude than the initial treatment.

We interpret the difference between the forced choice and the scale outcomes as evidence that blame is not necessarily zero-sum. While blaming the IMF might increase the degree by which voters hold the IMF responsible for reforms, it does not follow that they will reduce blame on other actors.

ATE/CACE: Blame Ruto (Scale)  ATE/CACE: Blame Ruto (Choice) ATE  ATE CACE  CACE

```text
−0.2  0.0  0.2  0.4  0.6  −0.2  0.0  0.2
```

Coefficient estimates and 95% confidence intervals  Coefficient estimates and 95% confidence intervals

ATE/CACE: Blame IMF (Scale)  ATE/CACE: Blame IMF (Choice)

ATE  ATE CACE  CACE

```text
−0.2  0.0  0.2  0.4  0.6  −0.2  0.0  0.2
```

Coefficient estimates and 95% confidence intervals  Coefficient estimates and 95% confidence intervals

Figure 6: IMF Blame Treatment and Respondent Blame Attribution: The four panels present both the ATE and CACE estimating blame for either President Ruto or the IMF in either a closed choice or scale format as a function of the IMF blame treatment. The dots indicate the point estimates and the bars indicate the 95% confidence intervals. Each model is estimated with HC2 robust standard errors.

### Discussion and Conclusion

Despite playing a large role in explaining why countries invite IMF intervention, evidence that scapegoating is effective is scarce. Existing efforts to test scapegoating arguments with observational data face methodological limitations. In this study, we tried to address these limitations by directly studying this phenomenon at the individual level and by randomly varying blame attribution for economic reforms in fictitious news reports. We then measured outcomes along the causal chain. Our treatment was successful at increasing blame toward the IMF. However, we find very small or null effects that IMF scapegoating increases incumbent approval or decreases incumbent blame for painful economic reforms.

Our study has several limitations that future research might want to address. Most

ATE/CACE: Blame Ruto (Scale)  ATE/CACE: Blame Ruto (Choice)

ATE  ATE CACE  CACE

```text
−0.2  0.0  0.2  0.4  0.6  −0.2  0.0  0.2
```

Coefficient estimates and 95% confidence intervals  Coefficient estimates and 95% confidence intervals

ATE/CACE: Blame IMF (Scale)  ATE/CACE: Blame IMF (Choice)

ATE  ATE CACE  CACE

```text
−0.2  0.0  0.2  0.4  0.6  −0.2  0.0  0.2
```

Coefficient estimates and 95% confidence intervals  Coefficient estimates and 95% confidence intervals

Figure 7: IMF Blame + Counter Narrative Treatment and Respondent Blame Attribution: The four panels present both the ATE and CACE estimating blame for either President Ruto or the IMF in either a closed choice or scale format as a function of the IMF blame + counter narrative treatment. The dots indicate the point estimates and the bars indicate the 95% confidence intervals. Each model is estimated with HC2 robust standard errors.

importantly, our findings are context dependent. Our study was conducted in Kenya under an existing IMF program, with the potential for greater IMF involvement. Further research might want to time the survey experiment with the roll-out of IMF-backed reforms. It is possible that the effectiveness of IMF scapegoating decays as time goes on and the public becomes more informed and attitudes potentially calcify.

The effect may also emerge in different political contexts. For example, we found little evidence of conditional treatment effects across supporters of the incumbent and opposition. However, the utility of IMF scapegoating may be depend on the identity of the incumbent. For obvious reasons, we could not randomize the incumbent in our study. However, future studies may seek to examine if IMF scapegoating works better when the incumbent has a more credible claim to oppose these reforms. In our case, Ruto did appear during campaigning as being an advocate for the working class given his ‘hustler-in-chief’ reputation. However, that image changed and he was bestowed with the nickname ‘Zakayo’ which is a local translation of the biblical character Zacchaeus, a chief tax collector from Jericho.

As with most survey experiments, our treatment was fleeting (Mutz, 2011). It remains plausible that repeated exposure to a IMF scapegoating frame may have a more sizeable effect on incumbent support (Lecheler et al., 2015).

Lastly, our research presents limited evidence that the opposition can counter IMF scapegoating. Future studies might consider examining the effect in environments where the opposition does not have the freedom or credibility to express blame. If our initial results are a guide, the scapegoat effect may have more mileage in non-competitive political environments.

### References

Abbott, Kenneth W. and Duncan Snidal (1998). Why states act through formal international organizations. The Journal of Conflict Resolution 42 (1), 3–32.

Abouharb, M Rodwan , Bernhard Reinsberg, and Tim Heinkelmann-Wild (2023). Does the messenger matter for blame-shifting to international organizations? evidence from the international monetary fund. EPSA XIII conference (Glasgow, 22-24 June).

Alca˜niz, I. and T. Hellwig (2011). Who’s to blame? the distribution of responsibility in developing democracies. British Journal of Political Science 41 (2), 389–411.

Auvinen, Juha Y (1996). Imf intervention and political protest in the third world: a conventional wisdom refined. Third World Quarterly 17 (3), 377–400.

Bird, G. and T. D. Willett (2004). Imf conditionality, implementation and the new political economy of ownership. Comparative Economic Studies 46 (3), 423–450.

Biten, Merve Theresa , Theresa Kuhn, and Wouter van der Brug (2023). How does fiscal austerity affect trust in the european union? analyzing the role of responsibility attribution. Journal of European Public Policy 30 (6), 1033–1050.

Blair, Graeme , Jasper Cooper, Alexander Coppock, and Macartan Humphreys (2019). Declaring and diagnosing research designs. American Political Science Review 113 (3), 838–859.

Bloniarz, Adam , Hanzhong Liu, Cun-Hui Zhang, Jasjeet S Sekhon, and Bin Yu (2016). Lasso adjustments of treatment effect estimates in randomized experiments. Proceedings of the National Academy of Sciences 113 (27), 7383–7390.

Bratton, Michael and Robert Mattes (2003). Support for economic reform? popular attitudes in southern africa. World Development 31 (2), 303–323.

Bratton, Michael , Robert Mattes, and Emmanuel Gyimah-Boadi (2005). Public opinion, democracy, and market reform in Africa. Cambridge University Press.

Chong, Dennis and James N Druckman (2010). Dynamic public opinion: Communication effects over time. American Political Science Review 104 (4), 663–680.

Clifford, Scott , Geoffrey Sheagley, and Spencer Piston (2021). Increasing precision without altering treatment effects: Repeated measures designs in survey experiments. American Political Science Review 115 (3), 1048–1065.

de Vries, Catherine E. , Sara B. Hobolt, and Stefanie Walter (2021). Politicizing international cooperation: The mass public, political entrepreneurs, and political opportunity structures. International Organization 75 (2), 306–332.

Dreher, Axel and Martin Gassebner (2012). Do imf and world bank programs induce government crises? an empirical analysis. International Organization 66 (2), 329–358.

Dreher, Axel , Valentin Lang, B. Peter Rosendorff, and James Raymond Vreeland (2018). Buying votes and international organizations: The dirty work-hypothesis. Available at SSRN 3278665 .

Dreher, Axel and Roland Vaubel (2004). Do imf and ibrd cause moral hazard and political business cycles? evidence from panel data. Open economies review 15 (1), 5–22.

Dreher, Axel and Stefanie Walter (2010). Does the imf help or hurt? the effect of imf programs on the likelihood and outcome of currency crises. World Development 38 (1), 1–18.

Duch, Raymond M. and Randy Stevenson (2006). Assessing the magnitude of the economic vote over time and across nations. Electoral Studies 25 (3), 528–547.

Edwards, S. and J. A. Santaella (1993). Devaluation controversies in the developing countries: Lessons from the bretton woods era. In M. D. Bordo and B. Eichengreen (Eds.), A Retrospective on the Bretton Woods System. Chicago: University of Chicago Press.

Gabel, Matthew and Kenneth Scheve (2007). Estimating the effect of elite communications on public opinion using instrumental variables. American Journal of Political Science 51 (4), 1013–1028.

Hays, Jude C. , Sean D. Ehrlich, and Clint Peinhardt (2005). Government spending and public support for trade in the oecd: An empirical test of the embedded liberalism thesis. International Organization 59 (2), 473–494.

Heinkelmann-Wild, Tim and Bernhard Zangl (2020). Multilevel blame games: Blame-shifting in the european union. Governance 33 (4), 953–969.

Hinterleitner, Markus and Fritz Sager (2017). Anticipatory and reactive forms of blame avoidance: Of foxes and lions. European Political Science Review 9 (4), 587–606.

Hobolt, Sara , James Tilley, and Susan Banducci (2013). Clarity of responsibility: How government cohesion conditions performance voting. European journal of political research 52 (2), 164–187.

Hobolt, Sara B. , James Tilley, and Jill Wittrock (2013). Listening to the government: How information shapes responsibility attributions. Political Behavior 35 (1), 153–174.

Hood, Christopher (2011). The Blame Game: Spin Bureaucracy and Self-Preservation in Government. Princeton University Press.

Jamal, Amaney and Helen V. Milner (2019). Economic self-interest information and trade policy preferences: Evidence from an experiment in tunisia. Review of International Political Economy 26 (4), 545–572.

Kalbhenn, Anna and Livio Stracca (2015). Does fiscal austerity affect public opinion? ECB Working Paper No. 1774 .

Kaufman, Robert R. and Leo Zuckermann (1998). Attitudes toward economic reform in mexico: The role of political orientations. American Political Science Review 92 (2), 359– 375.

Kern, Andreas , Elias Nosrati, Bernhard Reinsberg, and Dilek Sevinc (2023). Crash for cash: Offshore financial destinations and imf programs. European Journal of Political Economy 78, 102359.

Kharono, Brenda , Anne Kaggiah, Cyrus Mugo, David Seeh, Brandon L Guthrie, Megan Moreno, Grace John-Stewart, Irene Inwani, and Keshet Ronen (2022). Mobile technology access and use among youth in nairobi, kenya: implications for mobile health intervention design. Mhealth 8.

Konstantinidis, Nikitas and Bernhard Reinsberg (2023). Government ownership of imf conditionality programs: Conceptualization, measurement, and validation. International Studies Quarterly 67 (4), sqad074.

Kosmidis, S. (2018). International constraints and electoral decisions: Does the room to manoeuvre attenuate economic voting? American Journal of Political Science 62 (1), 519–534.

Kriegmair, Lisa , Berthold Rittberger, Bernhard Zangl, and Tim Heinkelmann-Wild (2022). Dolce far niente? non-compliance and blame avoidance in the eu. West European Politics 45 (5), 1153–1174.

Kumlin, Staffan (2011). Claiming blame and giving credit? unintended effects of how government and opposition frame the europeanization of welfare. European Union Politics 12 (4), 575–595.

Lecheler, Sophie , Mario Keer, Andreas RT Schuck, and Regula H¨anggli (2015). The effects of repetitive news framing on political opinions over time. Communication Monographs 82 (3), 339–358.

Lobo, Marina Costa and Rosella Pannico (2020). Increased economic salience or blurring of responsibility? economic voting during the great recession. Electoral Studies 65, 102141.

Louis, Melanie and Lucile Maertens (2021). Why International Organizations Hate Politics Depoliticizing the World. Routledge.

Maier, Michaela , Silke Adam, and J¨urgen Maier (2012). The impact of identity and economic cues on citizens’ eu support: an experimental study on the effects of party communication in the run-up to the 2009 european parliament elections. European Union Politics 13 (4), 580–603.

Margalit, Yotam (2013). Explaining social policy preferences: Evidence from the great recession. American Political Science Review 107 (1), 80–103.

Moravcsik, Andrew (1994). Why the european union strengthens the state: Domestic politics and international cooperation. Technical Report CES Working Paper No. 52, Harvard University.

Mutz, Diana C (2011). Population-based survey experiments. Princeton University Press.

Neundorf, Anja and Aykut ¨Ozt¨urk (2023). How to improve representativeness and costeffectiveness in samples recruited through meta: A comparison of advertisement tools. Plos one 18 (2), e0281243.

Ortiz, David G and Sergio B´ejar (2013). Participation in imf-sponsored economic programs and contentious collective action in latin america, 1980–2007. Conflict Management and Peace Science 30 (5), 492–515.

Pham, Katherine Hoffmann , Francesco Rampazzo, and Leah R Rosenzweig (2019). Online surveys and digital demography in the developing world: Facebook users in kenya. arXiv preprint arXiv:1910.03448 .

Reinsberg, Bernhard , Thomas Stubbs, and Louis Bujnoch (2023). Structural adjustment, alienation, and mass protest. Social Science Research 109, 102777.

Remmer, K. L. (1986). The politics of economic stabilization: Imf standby programs in latin america, 1954-1984. Comparative Politics 19 (1), 1–24.

Rickard, Stephanie J. (2015). Compensating the losers: An examination of congressional votes on trade adjustment assistance. International Interactions 41 (1), 46–60.

Rickard, Stephanie J and Teri L Caraway (2019). International demands for austerity: Examining the impact of the IMF on the public sector. The Review of International Organizations 14, 35–57.

Rittberger, Berthold , Hanna Schwarzenbeck, and Bernhard Zangl (2017). Where does the buck stop? responsibility attributions in complex international institutions. Journal of Common Market Studies 55 (4), 909–924.

Rosenzweig, Leah R and Yang-Yang Zhou (2021). Team and nation: Sports, nationalism, and attitudes toward refugees. Comparative Political Studies 54 (12), 2123–2154.

Schleiter, Petra and Margit Tavits (2018). Voter reactions to incumbent opportunism. The Journal of Politics 80 (4), 1183–1196.

Schlipphak, Bernd and Oliver Treib (2017). Playing the blame game on brussels: the domestic political effects of eu interventions against democratic backsliding. Journal of European Public Policy 24 (3), 352–365.

Smith, Alastair and James R Vreeland (2006). The survival of political leaders and imf programs: Testing the scapegoat hypothesis. In S. Kosack, G. Ranis, and J. Vreeland (Eds.), Globalization and the Nation State: The Impact of the IMF and the World Bank. London: Routledge.

Sommer, Moritz (2020). Blame-shifting in times of permanent austerity: Evidence from spain and greece in the eurozone crisis. JCMS: Journal of Common Market Studies 58 (3), 562–579.

Tallberg, Jonas (2002). Delegation to supranational institutions: Why, how, and with what consequences? West European Politics 25 (1), 23–46.

Thatcher, Mark and Alec Stone Sweet (2002). Theory and practice of delegation to nonmajoritarian institutions. West European Politics 25 (1), 1–22.

Traber, Daniel , Martijn Schoonvelde, and Gijs Schumacher (2020). Errors have been made, others will be blamed: Issue engagement and blame shifting in prime minister speeches during the economic crisis in europe. European Journal of Political Research 59 (1), 45–67.

Vaubel, R. (1986). A public choice approach to international organization. Public Choice 51, 39–57.

Vreeland, James Raymond (1999). The imf: lender of last resort or scapegoat. Yale University.

Vreeland, J. R. (2003). The IMF and economic development. Cambridge: Cambridge University Press.

Walter, Stefanie (2010). Globalization and the welfare state: Testing the microfoundations of the compensation hypothesis. International Studies Quarterly 54 (2), 403–426.

Weaver, R. Kent (1986). The politics of blame avoidance. Journal of Public Policy 6 (4), 371–398.

Williams, L. K. (2012). Pick your poison: economic crises, international monetary fund loans and leader survival. International Political Science Review 33 (2), 131–149.

### Appendix

#### Dependent Variables

Blame Attribution: To test if the treatments have manipulated blame for the austerity policies, we first ask questions regarding responsibility for economic reforms. These outcomes are also useful for exploratory research to examine if the treatments influence attribution of blame but do not extend to changes in overall presidential approval.

- “How responsible do you think President William Ruto is for the recent tax increases and reduction in fuel subsidies?” (11-point scale, Not responsible - Completely responsible)

- “How responsible do you think the International Monetary Fund (IMF) is for the recent tax increases and reduction in fuel subsidies?” (11-point scale, Not responsible - Completely responsible)

- “Who do you think is most responsible for the recent tax increases and reduction in fuel subsidies?” (The International Monetary Fund, William Ruto, Don’t know)

- “There have been massive protests in recent months against the policies to raise taxes and reduce fuel subsidies. These protests have often turned violent, with widespread destruction of property and looting. Who do you believe is most responsible for this violent unrest?” (President William Ruto, The International Monetary Fund, Opposition leader Raila Odinga, Other)

Presidential Approval Index:  Our central dependent variable is an additive index variable based on several questions aimed at capturing approval of President Ruto. These incumbent support questions are based on those constructed by Schleiter and Tavits (2018). The construction of an index aims to increase precision in estimation of individuals’ attitudes towards the incumbent. This index variable ranges from 0 (low approval) to 44 (high approval). A timer is set for each question so that respondents cannot navigate beyond the question for at least 4 seconds.

- “We will now ask you for your opinion about President William Ruto. We want to know what you think! Please take your time and read each question carefully.”

- “How would you rate the performance of President William Ruto?” 11-point scale from “Very negative” to “Very positive”

- “How would you rate the performance of President William Ruto in managing Kenya’s economy?” 11-point scale from “Very negative” to “Very positive”

- “Imagine that presidential elections will be held tomorrow. How likely would you be to vote for William Ruto?” 11-point scale from “Very unlikely” to “Very likely”

- “How well do you think President William Ruto will manage the Kenyan economy in the future?” 11-point scale from “Very badly” to “Very well”

Quasi-Behavioral Outcome: To examine if the treatments impact intentions beyond approval, we ask respondents about their willingness to sign a petition against the IMF and Government. We then take the difference between the petition responses to indicate how willing individuals are to signal disapproval of the IMF relative to the government.

- Would you be willing to sign a petition (at the end of this survey) asking the government to reverse the cuts to government spending? 11-point scale from “Very unwilling” to “Very willing”

- Would you be willing to sign a petition (at the end of this survey) asking the International Monetary Fund to reverse the cuts to government spending? 11-point scale from “Very unwilling” to “Very willing” Finally we ask if respondents would be willing to join a protest against the IMF.

- “How likely would you be to join a protest against the International Monetary Fund in the future?” 11-point scale from “Very unlikely” to “Very likely”

#### Post-treatment

- Randomized attention checks: respondents are randomly shown one of the following four questions: – “People are very busy these days and many don’t have time to find the best price for products. There are many websites that offer the same product at very different prices. Some have time to browse products all day, but others don’t even have time to read the questions carefully. To show that you’ve read this far, ignore the question below and just type the word ”green” in the text box below. How many websites do you visit to compare prices before finally purchasing a product?” – “We care about the quality of our survey data. It is very important for our research that we collect accurate measures of your opinion. We hope that you will provide honest, thoughtful answers to all questions in this survey. So far, have you provided honest, thoughtful answers to the questions in this survey?” (No, my answers have not been honest / I cannot say either way / Yes, all my answers have been honest / Yes, but only some of my answers have been honest) – “We are interested in how often citizens use financial services like online banking and in-person banking. Some people have multiple bank accounts, while others only have one bank account, or none at all. Some only use banking services in person, while others do all of their banking online or via their mobile phone. To show that you are paying attention, ignore the question below and just select the number eight. How many bank accounts do you have? ” – “Read the text below: The day began with a clear blue sky with white clouds. The group started on their journey and soon arrived in a small village. They saw that bright green grass surrounded a large house on the hill. Which colors are not mentioned?” (Red / White / Green / Blue / Yellow)

- Informational attention check: All respondents are asked a question to capture whether they were paying attention to the newspaper article treatments. Note: The first question below was asked to the first 100 respondents in the survey, collected during the soft-launch on 21 December 2023 conducted prior to the registration of this PAP. These responses indicated that the question wording was unclear and would not accurately capture differences in attention across treatment arms. As a result, this was replaced with the second question, which will be asked to the remaining 2,000 respondents. – (Attention check, soft-launch only) “You were asked to read an article about the Kenyan economy. What information was included in the article?” (The government will lower income taxes, the government wants to increase taxes and cut fuel subsidies, the IMF is pushing the government to increase taxes and cut fuel subsidies, the unemployment rate is increasing). – (Attention check) “You were asked to read a newspaper article about the Kenyan economy. What was the title of that article?” (Government announces decrease in public sector employment / IMF demands more tax increases and cuts to fuel subsidies / Government announces more tax increases and cuts to fuel subsidies, / Unemployment expected to increase in 2024 / IMF demands more reforms; opposition blames Ruto)

- “How did you access this survey? Please be honest! You will still receive Ksh. 100 airtime no matter which option you choose.” (Facebook advertisement / Received the link from a friend / Whatsapp advertisement / Other / Facebook post / Instagram advertisement)

- (Debrief) “Thank you for completing this survey. Your response has been recorded anonymously. The newspaper article that you were shown was fictional and was created by our research team only for the purposes of this survey. However, the information in this article was accurate and describes real changes in Kenyan economic policy. To thank you for your time, we would like to offer you Ksh. 100 of airtime. Enter your phone number below to receive your airtime within 24 hours, then click Next. If you do NOT want to receive airtime, click Next.”

#### Pre-treatment questions

Respondents are screened out of the survey if they answer “I do not agree” to the consent form (first question) or if they do not answer “Facebook advertisement” to the second question.

- (Consent) “Thank you for agreeing to take part in this study. This is a 7-minute survey about the Kenyan economy. Please answer all questions completely. You can only take this survey once. To thank you for your time, we will send you Ksh. 100 of airtime at the end of the survey. We will ask you for your phone number to transfer the airtime, but we will not contact you again, an will not keep any record of your phone number. The data collected in this survey are completely anonymous and will be used for academic research. Your participation in this study is totally voluntary and you can withdraw your response at any time. We will use the data to conduct statistical analysis and draw general conclusions. Anonymous data will only be shared with third parties upon publication of any article resulting from the project. No individual respondent will be identified. By clicking ”I agree” you indicate that you are at least 18 years old, have read and understand this statement, and agree to take part in this research survey.” (I agree, I do not agree)

- “How did you access this survey? Please answer honestly.” (Facebook advertisement, Instagram advertisement, Received the link from a friend, Other)

- “What is your current mobile service provider?” (Safaricom, Airtel Kenya, other)

- “How old are you?”

- “What is your gender?”

- “Which county do you live in?” (List of 47 counties)

- “What is your highest level of education?” (8-level scale from “No formal schooling” to “Post-graduate degree”)

- “What is the average monthly income of your household?” 8-point scale from “Less than Ksh. 1,200” to “More than Ksh. 30,000”

- “What is your opinion of President William Ruto?” 11-point scale from “Very negative” to “Very positive”

- “What is your opinion of opposition leader Raila Odinga?” 11-point scale from “Very negative” to “Very positive”

- “Have you ever heard of the International Monetary Fund (IMF) (Yes/No/not sure)”

- “What is your opinion of the International Monetary Fund (IMF)?” 11-point scale from “Very negative” to “Very positive”

- “Who did you vote for in the 2022 presidential election?” (William Ruto, Raila Odinga, Other, I did not vote / Prefer not to say)

- “Do you agree or disagree with the following statements?” [5-point scale strongly agree - strongly disagree] – “The government budget is too large and the government needs to cut spending” – “The government needs to cut spending to repay Kenya’s large debt” – “The government needs to increase taxes to repay Kenya’s large debt”

#### Estimation & Sampling Sampling

We will draw on a sample of 2,100 Kenyan adults recruited via a Facebook advertising campaign. Facebook users with accounts registered within Kenya will be shown an ad which invites them to participate in a short survey, in which they can earn mobile phone minutes (“airtime”) as compensation. The ad contains a link to the survey questionnaire, hosted in Qualtrics.

Social media recruitment has several advantages in the Kenyan context. A large share of the population has access to internet and has a social media account, with 98% of adults having access to a mobile phone (Kharono et al., 2022) and 49% with a Facebook account. While mobile phone and social media use skews towards wealthier, highly educated urban citizens, this is also a problem when recruiting online respondents from the panels of traditional survey companies in the region. Recruitment via Facebook allows access to a wide pool of respondents at a low cost. This sampling approach is similar to those previously used by Rosenzweig and Zhou (2021) and Pham et al. (2019) to recruit online survey respondents in sub-Saharan Africa.

First, a Facebook page was established for the organization that will deploy the advertisements. This is the “MiDebt Project,” which refers to the “Microfoundations of Debt Crises” research project under which this survey is being conducted. The page contains accurate information about the project as well as links to the project page on the Leiden University website. The Facebook ad will display the name “MiDebt Project” as the sponsoring organization, and users can navigate to the public MiDebt Project Facebook page directly from the ad.

Second, an advertisement campaign will be launched from the MiDebt Project page. The ad will be shown to Facebook accounts registered in Kenya to users above 18 years of age and will continue to be displayed to users until 2,100 complete survey responses have been collected. The ad will not be shown to a random sample of Kenyan Facebook users, but will be shown to an audience most likely to click on the link within the ad (as determined by the Facebook advertising auction system algorithm). Further, the ad will be targeted to be shown to separately to male and female users, and to users based on broad age categories derived from the results of the 2019 Kenyan census.[^10] This targeting approach follows the recommendations outlined in Neundorf and ¨Ozt¨urk (2023), where the authors find that maximizing link-clicks and targeting ads along broad demographic quotas like age and gender can improve representation in the sample.

Figure A1 shows the Facebook ad used for recruitment. This ad is shown to respondents either as a sponsored post within their timeline, or as a sidebar on the margins of a Facebook page.

Once users have clicked on the link, they are redirected to the survey questionnaire. Importantly, the Facebook advertising platform does not provide any granular data on which accounts that view or interact with ads. Thus, there is no way for the research team to connect survey respondents to their Facebook profile, or any personal information therein. Within the survey, respondents will first be asked a series of demographic questions (age, gender, region, income, education). We will set nationally representative quotas for age and gender, and screen out respondents after these quotas are filled. In line with previous survey research via social media in developing countries (Rosenzweig and Zhou, 2021; Pham et al., 2019), we expect our Facebook sample to be skewed towards urban, highly educated, and wealthy citizens and so will not attempt to achieve a nationally representative sample along these attributes.

Respondents are compensated for their participation with Ksh. 100 (€0.60) airtime minutes. This is delivered to them via text message. At the end of the survey, respondents are asked for their phone number in order to receive their airtime balance. Using the Africa’s Talking API (https://africastalking.com/), we will a send short-code text message to each phone number containing the airtime balance, and the message “Thank you for taking our survey! Here is your Ksh. 100 airtime.” This text message appears on respondents’ mobile phones as originating from the sender ID “LEIDEN POL” and respondents are unable to reply to this message. Immediately following the airtime transfer, respondents’ phone numbers are deleted (as respondents are informed in the survey consent form).

We use several strategies to ensure that individuals only take the survey once, and to guard against the survey link being redistributed outside of the Facebook advertising platform. First, we will activate Qualtrics fraud detection which attaches a cookie to the user’s browser, preventing the same device-browser combination from submitting subsequent survey responses. Second, we will discard responses that submit the same telephone number for compensation as has already been submitted in a previous response. Third, respondents

Figure A1: Facebook ad used for respondent recruitment

will be asked at the start of the survey where they accessed the survey link, and those who choose an answer other than “Facebook advertisement” will be screened out.

Data collection begins on 21 December and is expected to take 7 to 12 days. First, 100 responses were collected as a “soft-launch” to ensure that data collection and storage worked as expected. One small change was made following this soft launch: the replacement of one of the attention check questions (see page 14). The survey will be re-opened to collect the remaining 2,000 responses on 22 December following the registration of this PAP.

#### Pre-registered Exploratory Hypotheses

In addition to our primary hypotheses tested above, we consider several exploratory hypotheses that propose conditional relationships between blame attribution and incumbent support. Hypothesis 4 expects that the treatment will have a stronger effect on fiscal liberals than on fiscal conservatives. Voters who are ex ante fiscally liberal, meaning that they generally oppose spending cuts and tax increases, may strongly disapprove of the conservative fiscal policies needed to balance budgets and often implemented by the IMF. They may be the most likely to blame those responsible for the introduction of reforms, be that the government or the IMF, and so should respond most dramatically to information that identifies the responsible actor.

Hypothesis 4. Exploratory hypothesis The IMF-blame treatment will have a stronger effect among those who are fiscally liberal (opposed to spending cuts and tax increases).

The effect of blame-avoidance may also cleave along partisan lines. Governments may expect a loss of support when implementing structural adjustment, and so use scapegoating strategies to prevent supporters from defecting to the opposition. If this is effective, it would suggest that blame-avoidance has the strongest effect on the incumbent’s supporters. These voters may be most receptive to government attempts to frame structural adjustment as an externally-imposed necessity, and thus most likely to shift blame for reforms onto the IMF.

Hypothesis 5. Exploratory hypothesis The IMF-blame treatment will have a stronger effect among supporters of the incumbent.

However, we also consider that the opposite may be true, and that blame-avoidance may have the strongest effect on voters who support the opposition rather than the incumbent. Rather than scapegoating the IMF to retain its current supporters, the government may engage in blame-avoidance to avoid a deepening of the opposition, especially if opposition to reforms may lead to protest or unrest. Opposition supporters have low levels of support for the incumbent by definition, and so blame-avoidance framing is unlikely to suffer from the ceiling effects that may limit the effect of this strategy among the incumbent’s supporters.

Hypothesis 6. Exploratory hypothesis The IMF-blame treatment will have a stronger effect among supporters of the opposition.

#### Estimation

Parameters:

- Y a = Presidential Approval Index (0-11)

- Y b = Incumbent blame attribution (0-11)

- G = Government Blame (baseline condition)

- M = IMF blame (Treatment 1)

- C = IMF & government blame (Treatment 2)

- F = Fiscal liberalism

- I = Supports incumbent

- λD = matrix of pre-treatment covariates and their coefficients In each equation, we include a matrix of pre-treatment covariates and their coefficients on the right hand side λD. We select these variables agnostically with a LASSO selection model following the recommendation of Bloniarz et al. (2016). Essentially, We include each continuous variable individually and each categorical variable as dummies in a model predicting the outcome. The LASSO model returns only variables with non-zero coefficients.

We then include these predicted variables in our model estimating treatment effects.

Hypothesis 1. estimator: OLS, Yia = β0 + β1Mi + β2 Ci + λD + ϵi , expectation: β1 is positive and significant.

Hypothesis 2. estimator: OLS, Yia = β0 + β1Ci + β2Gi + λD + ϵi , expectation: β1 is negative and significant.

Hypothesis 3. estimator: OLS, Yib = β0 + β1 Mi + β2Ci + λD + ϵi , expectation: β1 is negative and significant.

Hypothesis 4. estimator: OLS, Yia = β0 + β1Mi + β2 Fi + β3(M * F )i + β4 Ci + λD + ϵi , expectation: β3 is positive and significant.

Hypothesis 5. estimator: OLS, Yia = β0 + β1Mi + β2 Ii + β3(M * I)i + β4Ci + λD + ϵi , expectation: β3 is positive and significant.

Hypothesis 6. estimator: OLS, Yia = β0 + β1Mi + β2 Ii + β3(M * I)i + β4Ci + λD + ϵi , expectation: β3 is negative and significant.

Beyond the Intention to treat effect (ITT) estimand in the above equations, we also look at the complier average causal effect (CACE) to address issues of attention. We estimate this in a two stage equation. To simplify things, we only estimate these equations on two experimental arms and omit observations for irrelevant arms (unlike above). In each tests (H1 and H2), we adopt the following two-staged least squares structure:

Complyi = γ0 + β1Ci + λX D + µi  (1) Yia = β0 + β1Complyi + λX D + ϵi  (2)

Figure A2: Minimum Detectable Effect (MDE)

#### Inference criteria

- We estimate our models with HC2 standard errors.

- For our primary hypotheses (1 & 2), we use a two-tailed test.

- we set α = 0.05 and will reject the null when the p-value is less than 0.05.

- for our remaining hypotheses estimating Y a, we adjust for multiple comparisons by running simulations and considering the correlation among the outcomes. See https://egap.org/resource/10-things-to-know-about-multiple-comparisons/ for more information on this approach.

#### Power analysis

This power analysis shows the statistical power at various assumed effect sizes (as a percentage of a standard deviation). This power analysis considers comparisons of two arms, just as our primary hypotheses, and thus we hold the sample size at 1400. The analysis in Figure A2 is based on 100 simulations of each effect size using the DeclareDesign Package in R (Blair et al., 2019). We assume a mean of 0 in the control group and a SD of 1 for all treatment arms. It also assumes a correlation of the outcome the additional covariates of R[^2] = 0.3. A total sample size of 2,100 will yield a minimum detectable effect of about 0.15 at 80% power.

### Notes

[^1]: A key exception is an article by Kosmidis (2018) that is related to scapegoating thought does not test blame attribution directly. It finds that treating citizens with narratives related to a government’s “room to maneuver” on economic policy has no effect on voting intentions.
[^2]: A timer is set for each vignette so that respondents are unable to navigate beyond the newspaper article for at least 10 seconds.
[^3]: A timer is set for each question so that respondents cannot navigate beyond the question for at least 4 seconds.
[^4]: We assumed that non-experimental covariates would correlate with the outcome at R[^2]=0.3.
[^5]: 40% 18-30 yrs, 24% 31-40 yrs, 16% 41-50 yrs, 20% ¿50 yrs, see https://knbs.or.ke/visualizations/?page_id=3126
[^6]: This condition was not preregistered but was deemed necessary after the first look of the survey. We discarded 503 responses and recollected the equal amount of respondents after resetting quotas.
[^7]: It is important to note that, given the small effect sizes, the study is not sufficiently powered to find a statistically significant difference between the treatment conditions as we initially intended in our preregistration.
[^8]: It isn’t surprising that 3 of the 4 coefficients are insignificant. The goal of the index was to increase power by reducing measurement error. Thus significance is more likely with the additive index than with each outcome estimated independently.
[^9]: We asked the independent scale questions first and then followed with the forced choice.
[^10]: 40% 18-30 yrs, 24% 31-40 yrs, 16% 41-50 yrs, 20% ¿50 yrs, see https://knbs.or.ke/visualizations/?page_id=3126