Beyond Instrumentalism

To understand whether coethnic bias manifests in judicial decision making, we begin by problematizing the dominant account of ethnic identity in the developing world. An influential literature in comparative politics adopts an instrumentalist framework to theorize why ascriptive group identities become salient and why agents of the state may privilege coethnics (e.g., Bates Reference Bates1974; Chandra Reference Chandra2007). Focusing on contexts in which voters condition vote choice on the receipt of material inducements (Van de Walle Reference Van de Walle2001), scholars in this tradition argue that political elites favor coethnics in the provision of public goods because doing so advances the interests of important in-group supporters (Kramon and Posner Reference Kramon and Posner2013).Footnote ³

To what extent do instrumentalist theories generalize to the judiciary? Kenyan judges are not elected but appointed by the President, usually on the advice of the Judicial Service Commission. Although coethnic favoritism plausibly plays a role in judicial appointments, it is unclear why such considerations would translate into everyday judicial decision making. This is particularly true of cases concerning individuals with limited means to exert pressure on the courts, such as low-income persons involved in petty crimes or disputes. These cases do not concern significant political players, nor do they have an overt political agenda; judges thus lack clear incentive to rule a certain way. That is to say, there is no clear strategic rationale for judges to privilege coethnic over non-coethnic defendants when adjudicating everyday disputes.

While instrumental motivations seem largely absent in quotidian cases, judges may still possess unconscious, implicit biases predisposing them to be more harsh or lenient toward certain groups (Redfield Reference Redfield2017). To date, the study of implicit biases in judicial decision making has focused on Western judiciaries (Cohen and Yang Reference Cohen and Yang2019) where it has been shown that ideology is a strong predictor of judicial outcomes (and ideology is strongly correlated with race and ethnicity) (Harris and Sen Reference Harris and Sen2019). However, judicial decision making in these contexts is not typically characterized as strategic—that is, motivated by the expectation of material reward. Such bias is seen instead as reflecting long-standing cleavages between majority and minority groups. For example, studies of U.S. courts sometimes frame the treatment of Black defendants by white judges against the broader history of Black subjugation and white supremacy (Clarke Reference Clarke2018). Similar approaches have been used to understand majority–minority dynamics in other countries, such as Jewish–Arab interactions in Israeli courts (Grossman et al. Reference Grossman, Weinstein, Gazal-Ayal and Pimentel2016).

However, from a noninstrumental perspective, judges should be susceptible to group-based attachments and antagonisms just like ordinary citizens. In particular, implicit biases may lead judges to assign positive regard toward in-group members and negative regard toward out-group members (Oyserman et al. Reference Oyserman, Kemmelmeier, Fryberg, Brosh and Hart-Johnson2003; Paluck and Green Reference Paluck and Green2009). This kind of unconscious, affective bias may subsequently inform cognitive elements that shape how judges assign blame and responsibility (e.g., Fiske and Pavelchak Reference Fiske, Pavelchak, Sorrentino and Higgins1986) or perceive the moral character of the accused (e.g., Alicke Reference Alicke2000; Nadler and McDonnell Reference Nadler and McDonnell2012), which can in turn shape how judges interpret cases and render judgment.

Mechanisms of Bias: In-Group Favoritism or Out-Group Derogation?

The preceding discussion has highlighted how social psychological factors can shape judicial bias absent instrumental incentives. But is judicial bias a manifestation of favoritism toward members of the in-group or hostility toward members of the out-group?

Theories of social identity and self-categorization suggest that coethnic bias is likely driven by in-group favoritism (Tajfel and Turner Reference Tajfel, Turner, Worchel and Austin1979; Turner and Reynolds Reference Turner, Reynolds, Brown and Gaertner2001). Self-categorization as an in-group member requires the “assimilation of the self to the in-group category prototype and enhanced similarity to other in-group members” (Hewstone, Rubin, and Willis Reference Hewstone, Rubin and Willis2002). Research in this tradition also posits that individuals have a tendency to assign positive valence (such as trust, esteem, and positive regard) to members of their in-group without conscious reflection (Brewer Reference Brewer1999; Levin and Sidanius Reference Levin and Sidanius1999; Otten and Wentura Reference Otten and Wentura2001). Alternatively, it might be the case that coethnic bias is driven by out-group derogation. While group identification does not always lead to feelings of hostility toward outsiders, ample evidence suggests that such sentiments can be easily triggered in polarized societies that have a history of intergroup conflict (Stephan and Stephan Reference Stephan, Stephan and Oskamp2013). In such settings, members of the out-group are more likely to be perceived as a threat to the in-group, which can arouse feelings of fear, disgust, and thus antagonism toward the source of the threat (Sherif and Sherif Reference Sherif and Sherif1953). Whether judicial bias is driven by in-group positivity or out-group negativity generates different implications for our expectations about the affective content of written judgments, an issue we turn to in the next section.

Implicit Bias, Legal Outcomes, and Legal Writing

Recent work on judicial bias has examined specific types of stereotypes, including race and gender (e.g., Ash, Chen, and Ornaghi Reference Ash, Chen and Ornaghi2020; Rice, Rhodes, and Nteta Reference Rice, Rhodes and Nteta2019). In contrast to these approaches, our goal is to uncover the affective correlates of in-group favoritism and out-group derogation. To this end, we join a growing trend in the social sciences using text as data to trace difficult to measure concepts like sentiment and personality (Gennaro and Ash Reference Gennaro and Ash2021; Osnabrugge, Hobolt, and Rodon Reference Osnabrugge, Hobolt and Rodon2021; Ramey et al. Reference Ramey, Hollibaugh, Klingler and Hall2019).

A standard measure of judicial bias treats case outcomes as a discrete variable: whether an appeal is allowed or denied. We examine to what extent personal characteristics of the defendant (rather than legal matters of the case) affect how judges rule. However, attributing bias to either in-group or out-group attitudes is challenging if we only look at final verdicts, which lack context for understanding their motivation. That is, without a neutral control condition, we might observe that differential treatment exists, leaving unknown the motivation for that difference (Gazal-Ayal and Sulitzeanu-Kenan Reference Gazal-Ayal2010; Gill, Kagan, and Marouf Reference Gill, Kagan and Marouf2017; Harris and Sen Reference Harris and Sen2019).

However, the full text of a legal decision may provide more analytical leverage on these mechanisms. Before delivering a verdict, judges summarize each case and explain their decision’s logic. A written judgment can be conceived of as the final output of cognitive processes in which a judge uses evidence, legal concepts, and judicial discretion to support their decision (Maroney Reference Maroney2016; Rachlinski and Wistrich Reference Rachlinski and Wistrich2017; Simon Reference Simon1998).Footnote ⁴ Our research builds on a robust literature examining how affective framing and cues shape legal reasoning and outcomes (Beattey, Matsuura, and Jeglic Reference Beattey, Matsuura and Jeglic2014; Black et al. Reference Black, Treul, Johnson and Goldman2011; Liu and Li Reference Liu and Li2019; Wistrich, Rachlinski, and Guthrie Reference Wistrich, Rachlinski and Guthrie2015). Using this lens, we posit that if judges are indeed making decisions based on their implicit biases, such sentiments are likely reflected in their written legal judgments.

To dissect judgments for evidence of in-group versus out-group bias, we first identify what kinds of affective content we would expect to see if either mechanism were in play. Existing research finds in-group favoritism to be strongly associated with notions of trust; in the African context, such studies tend to portray trust as a basic behavioral regularity in coethnic interactions (e.g., Arriola, Choi, and Gichohi Reference Arriola, Choi and Gichohi2021; Robinson Reference Robinson2020). Social psychology research similarly argues that trust underpins in-group favoritism (Allport Reference Allport1954; Brewer Reference Brewer1999), whereas disgust more often accompanies feelings of out-group derogation (Hodson et al. Reference Hodson, Choma, Boisvert, Hafer, MacInnis and Costello2013; Mackie, Devos, and Smith Reference Mackie, Devos and Smith2000). From these findings, it is reasonable to expect that the presence of trust-related and disgust-related language in appeals judgments would imply biases suggesting in-group favoritism and out-group derogation, respectively.Footnote ⁵

Are Certain Groups More Predisposed to Bias?

Our main claim is that psychological mechanisms can predispose judges to demonstrate group-based bias in judicial decision making. However, this does not necessarily mean that judges of different groups are equally susceptible to such biases. Literature in social psychology and sociology suggest that prevalence of biases among groups corresponds to hierarchical status (e.g., Hagendoorn Reference Hagendoorn1995; Mullen, Brown, and Smith Reference Mullen, Brown and Smith1992). Two prominent theoretical strains are social identity and dominance theories and realistic group conflict theory.

Research on social identity theory and social dominance orientation predicts that processes of social comparison and social identification may lead members of higher-status groups to be more likely to discriminate between the in-group and out-group(s) (Sidanius et al. Reference Sidanius, Levin, Liu and Pratto2000; Tajfel and Turner Reference Tajfel, Turner, Worchel and Austin1979). “Dominant” groups derive esteem from their superior status, reinforcing the value and worth they attach to their dominant position (Sachdev and Bourhis Reference Sachdev and Bourhis1987). Being at the top of the social hierarchy makes such groups more predisposed to preserving the status quo as a means of sustaining their privileged access to resources and power (Harkness Reference Harkness2018; Sidanius and Pratto Reference Sidanius and Pratto2001). Evidence reveals higher levels of in-group bias among dominant group members in hierarchical societies in contexts as diverse as Israel, India, the Netherlands, Northern Ireland, and the United States (Levin Reference Levin2004; Levin and Sidanius Reference Levin and Sidanius1999).Footnote ⁶

In the Kenyan case, this framework implies that bias should be concentrated among judges belonging to the “dominant” ethnic group: the Kikuyu. As the numerically largest ethnic group in Kenya, Kikuyus have seen their leaders occupy the presidency three times (Jomo Kenyatta, Mwai Kibaki, and Uhuru Kenyatta), comprising most of the postindependence period. Their political dominance has spread to other branches of government, including the judiciary, with multiple Kikuyu jurists serving as Chief Justice.Footnote ⁷

By contrast, realistic group conflict theory in social psychology and political science argue that bias and discrimination mirror political conflicts among competing groups (Horowitz Reference Horowitz2000; Sambanis and Shayo Reference Sambanis and Shayo2013; Sherif and Sherif Reference Sherif and Sherif1953); intergroup animosity is rooted in competition over scarce resources (Gurr Reference Gurr2015). Out-group discrimination and conflict thus reflect existing grievances over the distribution of material goods (Sherif Reference Sherif1988). Other work in this tradition contends that intergroup cleavages result from the subordination of certain groups, entrenching intergroup animosity and systems of marginalization and exclusion (Wimmer, Cederman, and Min Reference Wimmer, Cederman and Min2009).

In the Kenyan context, this second perspective would lead us to expect bias to arise between specific constellations of ethnic groups. The Kikuyu and Kalenjin ethnic groups have long monopolized political power and engendered grievances among excluded groups, notably the Luo (Widner Reference Widner1992). The Kikuyu–Luo rivalry stems from the early independence struggle for national control between Kikuyu President Jomo Kenyatta and Luo Vice President Oginga Odinga.Footnote ⁸ Kalenjin–Luo relations have also been strained for decades, reflecting similar national-level power struggles.Footnote ⁹ More recently, Kikuyu–Kalenjin hostilities have intensified since the return to multiparty politics in the 1990s, culminating in the infamous postelection violence of 2007–08 that led to more than 1,000 dead and more than 600,000 displaced (Lynch Reference Lynch2014). Given Kenya’s long history of ethnic rivalries at every level of government from the presidency downward, these tensions might be on display in the courtroom.

Observable Implications

The preceding discussion suggests three observable implications. First, we expect that judges will exhibit identity-based implicit bias in their decisions, even absent strategic or instrumental considerations. Second, these biases may align with coethnicity and be driven by in-group favoritism; alternatively, out-group derogation may drive negative cross-ethnic biases. Patterns of trust- or disgust-related language in written judgments should thus mirror in-group or out-group motivated biases. Third, such biases might be higher among judges who belong to the dominant ethnic group; alternatively, bias may follow historical patterns of cross-ethnic rivalry.

Measuring Sentiment in Legal Judgments

We expect in-group favoritism or out-group derogation to motivate bias when judges evaluate an appeal. To test this, we use text-as-data approaches to assess the degree to which emotive reasoning appears in judicial writing. Our analysis builds on conventional dictionary methods wherein the count or proportion of key words in a given document is used to determine that document’s category (Grimmer and Stewart Reference Grimmer and Stewart2013). To this end, we generated word lists capturing our main mechanisms of interest: in-group favoritism or out-group derogation. Specifically, we identified terms related to trust and disgust to measure in-group favoritism and out-group derogation in written legal decisions.Footnote ¹⁷ We then calculated the number of trust and disgust terms as a proportion of total terms for each decision.Footnote ¹⁸

The technical nature of judicial writing makes this approach a challenging test of our proposition. Not surprisingly, official guidelines from the Kenya Criminal Procedure Benchbook explicitly discourage judges from invoking emotive sentiments in their decisions to avoid allegations of bias: “judgment should not contain derogatory language … a dispassionate approach and clear finding of fact, are more indicative of judicial approach, and do not lay the magistrate open to a charge of possible bias. The court may express strong condemnation of the conduct of the accused, but it must be careful not to be abusive or, for example, imply that the conduct is what might be expected of those belonging to a particular race, religion, etc.”Footnote ¹⁹

Although these rules are considered standard practice, the inclusion of explicit instructions suggests a broader concern among legal practitioners in Kenya—that judges may indeed be discriminatory in their legal opinions and need to guard against such tendencies when rendering verdicts. Given these cautions, we expect that any expression of emotion in a written judgment will likely reflect subtle, implicit, and often unconscious biases rather than overt prejudice.

Examples from our appeals corpus corroborate these expectations, suggesting that terms of trust and disgust are subtly expressed in legal writing. In one successful appeal, the judge described the appellant’s standing using terms of trust, remarking that the “magistrate erred in law and fact in disregarding the appellant’s defence, which was consistent and trustworthy[.]”Footnote ²⁰

In another ultimately denied judgment, the judge invoked terms of disgust in assessing the facts of the case: “the appellant had converted [the witness] into his wife, a shameful act indeed. She also physically suffered by the damage of her womanhood. The best description that this court can accord to the behavior of the appellant was that he was a beast to [the witness]. As rightly noted by the trial magistrate, he ought to be kept away from the society.”Footnote ²¹ Terms of disgust written into judgments reveal a judge’s personal and moral assessments of appellant character: “the offence committed was barbaric, immoral and had definitely left the complainant traumatized. I find that was a justification for passing sentence higher than the minimum provided and did not in any way offend the provisions of … the Constitution of Kenya[.]”Footnote ²²

These excerpts illustrate that trust and disgust dictionaries capture personal evaluations of appellants, distinct from an appeal’s legal merits. Given the technicality of legal writing, including the instruction to minimize perceptions of bias in the record, our approach is a hard test of the hypothesis that judicial bias is motivated by in-group favoritism or out-group derogation. If we find evidence that judges invoke terms of trust (disgust) with respect to coethnic (non-coethnic) appellants, we take it as consistent with the posited affective-cognitive mechanisms that may influence outcomes.

Research Design

To identify the effect of identity on criminal appellate judgments, we exploit features of the case assignment process in the Kenyan Courts. Cases arriving on the docket of each high court station are sorted into categories for assignment to legal divisions within the court station: family, commercial and admiralty, constitutional, land and environment, and, most relevant for our purposes, criminal. The deputy registrar, responsible for case scheduling, assigns each new incoming case to a judge. This intradivision assignment is determined by judges’ calendars and existing workloads, not by case characteristics: case assignment criteria—a judge’s schedule and case load—are orthogonal to case particulars like the ethnicity of defendants and judges. This provides us with quasi-random variation in the ethnic relationship between the appellant and the judge.

We use a linear model to examine the relationship between coethnicity and case outcomes:

(1) $$ {\mathrm{Y}}_{\mathrm{i}}={\unicode{x03B2}}_0+{\unicode{x03B2}}_1{\mathrm{M}}_{\mathrm{i}}+{\unicode{x03B2}}_2{\mathrm{X}}_{\mathrm{c}}+{\unicode{x025B}}_{\mathrm{i}}, $$

where Y_i is either the binary indicator for whether the appellate judge ruled in favor of the defendant or a measure of sentiment in text; M_i takes the value of one if the case was assigned to a coethnic judge of the defendant; and X_c is a vector of controls and fixed effects including courthouse-year, judge ethnicity, individual judge, and in our most restrictive specification, courthouse-year and individual judge fixed effects. We present our baseline results with our coethnic match variable and courthouse-year fixed effects and progressively add more restrictive sets of fixed effects and controls. Then, we conduct ethnic subgroup analyses by judge ethnicity to probe for differential bias.

Case allocation is a manual process, and it is possible that the principles of case assignments are not respected. Appendix Table B1 provides balance checks suggesting that the case assignment mechanism likely induced quasi-random variation in judge–defendant coethnicity. Conditional on the courthouse-year in which the case was heard, there is balance on most covariates. Most differences remain insignificant, except for the proportion of crimes classified as murder, manslaughter, or theft. Given this, we include regression specifications with these covariates, which do not alter our findings.

Identifying Mechanisms of Coethnic Bias: Favoritism or Derogation?

To elucidate whether in-group favoritism or out-group derogation drives coethnic bias in judicial decision making, we examine the texts of appeals judgments. We calculate the proportion of trust, disgust, positive, and negative sentiments for each written judgment using word embeddings. We estimate the effect of coethnicity using OLS as described in Equation 1 above.

Table 4 shows that only trust has a positive and significant relationship with the coethnic match variable. Judgments for coethnics contain approximately 0.12 standard deviations more trust-related words than judgments written for non-coethnics. A judge invokes terms of confidence, credibility, and honesty more often when writing a judgment for a coethnic. The insignificance for the disgust score suggests that judges use similar amounts of disgust-related language when writing for coethnics and non-coethnics. We interpret this as evidence that judicial bias is a manifestation of in-group favoritism rather than out-group derogation.

Table 4. Coethnic Bias in Written Judgments: Corpus Seeds, GloVe Vectors

Note: Coefficients estimated using OLS. “Coethnic match” is a binary variable equal to one if the judge and appellant share the same ethnic group, zero otherwise. *p < 0.1, ^** p < 0.05, ^*** p < 0.01.

We draw two main conclusions from this dictionary analysis, both of which should be interpreted with caution. First, despite the fact that judges are explicitly instructed to refrain from issuing decisions based on nonlegal considerations, they are still likely to invoke emotive sentiments (trust) when regarding coethnic appellants. This has important implications for legal reform efforts in Kenya, the bulk of which have either implicitly or explicitly assumed that judicial bias falls along gender, income, or other socioeconomic dimensions. Our findings suggest that more consideration should be directed instead toward the conscious or subconscious favoritism judges may harbor for coethnics. Second, our analysis reveals the utility of text-as-data approaches on judicial writing, especially the applicability of minimally supervised dictionary methods on niche corpora. Our main findings suggest that conventional dictionary methods combined with word embedding models elucidate different sources of implicit, affective bias in legal language and help assess the weight of one type of bias against another.Footnote ²⁶ Although these findings are more suggestive than definitive, they provide a framework for understanding some nuances of judicial decision making.