George Floyd’s murder renewed attention to police bias with far-reaching influence. Protests unfolded throughout the world, starting with a focus on the Black experience of racial discrimination, and extending to other minorities. For instance, Black Lives Matter, which emerged to protest anti-Black violence, sparked the sister movement Dalit Lives Matter. Dalits are considered the lowest-level caste in Hindu society, and have historically been subjected to “untouchability.”Footnote 1 Activists have called on governments and international organizations to combat caste-based discrimination and violence from law enforcement agencies, underscoring concerns around discriminatory behavior affecting people’s access to justice worldwide (HRW 2020). Others have challenged the existence of police bias, responding with countermovements like Blue Lives Matter. Against this backdrop, we examine whether the police discriminate with investigation decisions in Nepal.
Employing a conjoint survey with police officers that is representative of five districts in Nepal, we find evidence of caste-class discrimination. Officers are more likely to prefer targeting offenders from caste-class subjugated (CCS) communitiesFootnote 2 over those from privileged backgrounds. Specifically, we find that the police are more likely to personally prefer investigating low-caste Dalit offenders than high-caste ones, poor offenders than wealthy and middle-class ones, and illiterate offenders than literate ones. We also find that officers’ assessments of institutional investigatory preferences generally reveal the belief that the police are more likely to investigate low-caste Dalit offenders over high-caste Brahmin ones. However, we do not detect evidence of officers believing that the institution would target offenders based on class considerations alone. In sum, offenders from CCS communities have a higher probability of being targeted for investigation than Nepali society’s upper echelons.
Our contributions are fourfold. First, we employ a different method for detecting bias than past policing scholarship; while much of this work relies on administrative data, laboratory simulations, and qualitative approaches, we instead use a conjoint experiment.Footnote 3 Second, rather than studying the use of force or police stops, on which most of the literature focuses, we extend the debate on whether police discriminate by examining a more mundane decision: whether police prioritize investigations given offenders’ ascriptive characteristics. This has implications for access to justice: if police officers respond differently to an offender based on attributes of their identity rather than the crime in which they engage, then offenders and their victims are poised to receive different treatment from the legal system for reasons unrelated to the crime committed.
Third, our examination of discrimination incorporates considerations of institutional discrimination (Small and Pager Reference Small and Pager2020), thus enriching a literature that focuses overwhelmingly on the tension between two individual-level types of discrimination: statistical and taste-based. In doing so, we are able to see that individual officers do not perceive some of their own biases as existing collectively at the level of their institution. Finally, we extend the scholarly focus on policing beyond the United States, adding greater context validity—a component of external validity (Egami and Hartman Reference Egami and Hartman2022) referring to the transportability of treatment effects across contexts and settings (Bareinboim and Pearl Reference Bareinboim and Pearl2016). Because the literature is U.S.-centric, the issue of police discrimination is largely examined with respect to racial discrimination. Yet, dimensions of power do not always involve race. In many parts of South Asia, caste—a Hindu hereditary social stratification system—determines social hierarchy. By identifying the presence of caste-based police discrimination in Nepal, we extend empirical analysis of police bias beyond the U.S. border, with possible implications for the many countries (including India, Sri Lanka, Bangladesh, and Pakistan) and hundreds of millions of people living in societies that have historically been organized around a caste system.
ADDING TO SCHOLARSHIP ON POLICING BIAS
Numerous studies examine policing bias, with most focusing on racial biases. Although many find evidence of bias, the presence of and degree to which bias is detected depend on methodological techniques, samples, and outcome measures. Some studies suggest an absence of bias, or even a “counter-bias,” and have become highly visible in the political sphere, underscoring a lack of consensus on the existence and nature of police bias. Our systematic literature review reveals that 42% of studies report evidence of police bias, 33% report mixed findings, 21% report no evidence of bias, and 3% indicate evidence of anti-white bias (see Table A3 in Appendix C of the Supplementary Material). While our study does not adjudicate between these findings, given the mixed results, more scholarship is necessary to gain greater clarity on the magnitude and nature of policing bias.
Studies relying on administrative police records have fueled vigorous debate on policing bias, highlighting the challenges of detecting bias with administrative data. Fryer (Reference Fryer2019), for instance, did not find evidence of racial bias in use of lethal force. However, administrative data are often incomplete, as police records exclude information on civilians whom officers observe but do not investigate. If police discriminate when choosing whom to initially investigate, as our study suggests, using administrative records to estimate bias masks racially biased policing (Knox, Lowe, and Mummolo Reference Knox, Lowe and Mummolo2020). Indeed, police in South Asia have systematically underreported caste-based crimes by classifying homicides as suicides and denying Dalits entry into police stations (Narula Reference Narula2008). Other studies have examined ascriptive bias in policing using experimental methods including as-if random designs (e.g., Grogger and Ridgeway Reference Grogger and Ridgeway2006) or laboratory simulations (e.g., Mekawi and Bresin Reference Mekawi and Bresin2015). We add to the literature with a method of detecting bias that, to the best of our knowledge, has not previously been used to study police officers: a conjoint experiment.
Our research also focuses on a less scrutinized form of police behavior. While policing research overwhelmingly focuses on use of force or police stops as outcome measures (e.g., Cano Reference Cano2010), our study broadens the range of encounter scenarios to examine decisions about what police investigations to initially pursue. These other forms of police behaviors are important to consider when seeking to better understand the downstream effects that police bias may have on accessing justice.
Additionally, beyond examining the presence of policing bias, more scholarship focused on the mechanisms through which bias occurs is needed. Economic theories of taste-based discrimination (a “taste” for discriminating against a certain group) and statistical discrimination (rationally targeting certain groups that are statistically more likely to commit crimes), which focus on the root of discriminatory practices of individuals, dominate these mechanism discussions. Some scholars have recently criticized the validity of the theory of statistical discrimination, arguing that the theory itself reinforces stereotypes which then justify discriminatory behavior (Tilcsik Reference Tilcsik2021) and ignores “history, laws, and social norms” when determining an ascriptive characteristic such as race or caste as a meaningful indicator (Spriggs Reference Spriggs2020, 2). Instead, sociologists propose also considering institutional discrimination, a mechanism by which an institution can discriminate regardless of the biases of each of its members (Small and Pager Reference Small and Pager2020). We build on this literature by examining both personal preferences and officers’ assessment of institutional biases. Indeed, we find differences between officers’ personal preferences and their assessment of their institution’s preferences, underscoring the utility of looking beyond individual-level mechanisms.
Moreover, the literature faces limitations with respect to context validity (Egami and Hartman Reference Egami and Hartman2022). Most studies on police discrimination focus on racial biases and data from one police department or city primarily in the United States—an extremely decentralized system with over 18,000 different policing agencies (Banks et al. Reference Banks, Hendrix, Hickman and Kyckelhahn2016). It is thus difficult to generalize beyond a few American cities. Our study takes place in South Asia and, therefore, provides an opportunity to study policing bias outside of the United States. Moreover, by examining a centralized police agency where officers are regularly transferred between stations, with a response rate of nearly 100%, our findings likely extend beyond the interviewed officers.
Furthermore, our study expands the research on policing bias beyond a discussion of race-based inequality by considering caste-based discrimination. Ascriptive or status inequality refers to the unequal distribution of privilege derived from hierarchical social orders that are typically found in settings with histories of slavery, colonialism, and caste systems (Suryanarayan Reference Suryanarayan2019). The global similarities of ascriptive inequality, regardless of the causal basis of the inequality, come through well in Caste: The Origins of Our Discontents (Wilkerson Reference Wilkerson2020, 17), which draws parallels between the caste systems of India, race in the United States, and Nazi Germany, noting that “[e]ach version relied on stigmatizing those deemed inferior to justify the dehumanization necessary to keep the lowest-ranked people at the bottom and to rationalize the protocols of enforcement.” Finding empirical evidence of caste-based police bias, which is consonant with past findings of race-based police bias, we highlight the similar perceptions that police officers have toward individuals deemed inferior in these systems of ascriptive inequality, regardless of whether they are race- or caste-based.
Of course, discrimination on the basis of ascriptive characteristics is intertwined with a non-ascriptive characteristic: economic inequality. Soss and Weaver (Reference Soss and Weaver2017, 567) make this observation when discussing American policing: “race-class subjugated communities are positioned at the intersection of race and class systems, and these two dimensions of power relations remain thoroughly entwined in experiences of civic ostracism, social and political oppression, economic marginalization, and state-led governance.” This conception remains helpful when talking about policing issues globally. Indeed, low-caste Dalits in Nepal are poorer, on average, than upper-caste individuals (Banerjee and Knight Reference Banerjee and Knight1985; Jodhka and Shah Reference Jodhka and Shah2010). As such, we adopt Soss and Weaver’s approach by examining the interaction between socioeconomic class and caste, and adapt their terminology accordingly by using the term “caste-class subjugated” communities.
EMPIRICAL APPROACH
To evaluate police bias against CCS communities, we surveyed 1,065 Nepali police officers from five districts in Nepal’s most populous province, Bagmati Pradesh, in 2017: Bhaktapur, Chitwan, Dhading, Kavrepalanchowk, and Makwanpur.Footnote 4 Although we studied a sample that is representative of these districts,Footnote 5 our results also likely speak to the broader Nepali police force as the country’s police agency was centralizedFootnote 6 and required that officers be regularly transferred to new stations.Footnote 7 As expected with this requirement, survey results showed that officers came from 66 of Nepal’s 77 districts. By securing authorization from the Inspector General of Police, we achieved a response rate of 97.17%.
The sample consists of nearly 80% lower-ranking officers (head constables and constables); 20% mid-rank officers (inspectors, subinspectors, and assistant subinspectors); and 1% high-rank officers (district superintendents and deputy superintendents). Most participants are Hindu (93%) and about 43% are upper-caste Brahmin and Chhetri. They also constitute the majority of the high-rank (77%) and mid-rank (58%) positions in our sample. Comprising 6% of the sample, Newars similarly hold a disproportionate share of high-rank (15%) and mid-rank (7%) positions. Dalits constitute 5% of the sample and hold 6% of low-rank positions, 2% of mid-rank ones, and no superintendent positions (see Table A5 and Figure A2 in Appendix D of the Supplementary Material). We over-sampled women, who comprise 9% of officers in these districts, but nearly 16% of our sample. We thus consider sampling weights. The weighted average years of employment for surveyed officers is 12 and the weighted average monthly income is NRS 35,378 (USD 345).Footnote 8 The majority (63%) of respondents achieved an education of School Leaving Certificate (Class 11) or higher.Footnote 9
We employed a conjoint design to study police attitudes toward offender profiles across a range of crimes, holding presumed guilt constant.Footnote 10 This method allows us to causally examine the importance of attributes of interest relative to each other (Hainmueller and Hopkins Reference Hainmueller and Hopkins2015). Respondents reviewed pairs of profiles for hypothetical offenders (Appendix A of the Supplementary Material includes the exact script). Guilt of the profiled individuals was strongly implied, as each profile noted that the individual “engaged in” a particular crime. Each individual was randomly assigned characteristics from six attributes (see Table 1): caste/ethnicity, household income, education, gender, age, and crime type. This paper focuses on the first three attributes. Values for each attribute had equal probability of being randomly drawn and no restrictions were imposed (for profile display frequencies, see Figure A4 in Appendix F of the Supplementary Material).Footnote 11 These demographic attributes are typical information that officers might learn when investigating accused offenders. Research on caste in South Asia often examines differences between upper-caste Brahmins and Dalits. We include Dalit and Brahmin as values for caste/ethnicity along with Newar, an indigenous group primarily concentrated in our study sites that has high social and economic standing. While there are three other varnas within the Hindu caste system (Kshatriyas, Vaishyas, and the Shudras) and over 1,000 ethnic groups (Central Bureau of Statistics 2014), given statistical power considerations, we included just three groups, and chose them for their distinct positions within Nepal’s caste/ethnic hierarchy.Footnote 12
Attributes for Offender Profile in Conjoint Design
Note: Attribute values in Nepali can be found in Table A1 in Appendix A of the Supplementary Material.
After reviewing the paired profiles, officers responded to choice-based outcome questions. We asked officers, “If you had to choose between them, which of these two individuals would you personally prefer to investigate?” This “forced-choice” design encourages respondents to consider trade-offs when choosing which individual to investigate. We also asked respondents, “If you had to choose between them, which of these two individuals do you think the police would investigate?” We coded responses to both questions as dichotomous variables, where 1 denotes the preferred individual, and 0 otherwise. Respondents completed this process twice: 1,065 police officers collectively evaluated 2,130 pairings, for a total of 4,260 profiles.
We estimated the average marginal component effect (AMCE) following Hainmueller, Hopkins, and Yamamoto (Reference Hainmueller, Hopkins and Yamamoto2014). The AMCE represents the marginal effect of a specific attribute averaged over the joint distribution of the remaining attributes. For our first outcome of interest, for example, the point estimates of each attribute indicate the marginal effect on the probability that the offender profile is personally preferred for investigation. We used generalized linear model (GLM) regression to estimate the AMCE with robust standard errors correcting for within-respondent clustering.Footnote 13
RESULTS
Figure 1 displays the estimates of the marginal effects of each attribute on our two outcomes: the probability officers personally prefer to investigate an offender (Figure 1a) and the probability officers actually believe that the police will investigate the offender (Figure 1b). Attribute values are listed along with their corresponding point estimates.Footnote 14 As shown in Figure 1a, across education, income, and caste/ethnicity, police personally prefer to investigate offenders from more disadvantaged backgrounds than those from more advantaged ones. Specifically, compared with Dalits, Newars are 3.7 percentage points (pp) less likely to be preferred for investigation (
$ p=0.044 $
) and Brahmins are 2pp less likely to be preferred, although this effect is not statistically significant (
$ p=0.268 $
). It is perhaps not surprising that the police view Newars more like Brahmins than Dalits. Newars are an indigenous ethnic group with its own caste system (Gellner Reference Gellner, Gellner and Quigley1995) and many hold a privileged position economically, politically, and socially (Gurung et al. Reference Gurung, Suwal, Pradhan and Lama-Tamang2014). They have the highest per capita income of any caste/ethnic group according to the 2011 Nepal Living Standards Survey.
Effects of Offender Attributes on Police Investigatory Preferences
Note: Estimates are based on the GLM model with clustered standard errors at the respondent level. The bars indicate 95% confidence intervals and the points without bars denote the reference category. Point estimates in black convey statistically significant effects. Estimates of panels (a) and (b) of Figure 1 correspond to coefficients in columns 1 and 5 of Table A6 in Appendix F of the Supplementary Material, respectively.
Pertaining to class, offenders from middle-income and wealthy households are, respectively, 4.1pp (
$ p=0.021 $
) and 4.7pp (
$ p<0.01 $
) less likely to be preferred for investigation than offenders from poor households. Similarly, literate offenders are 4.2pp (
$ p<0.01 $
) less likely to be preferred over illiterate offenders. Gender is not the focus of this study, but notably, police officers are less likely to prefer investigating men rather than women by 4.6pp (
$ p<0.01 $
).Footnote 15
Moving beyond officers’ personal investigatory preferences, we examine their perceptions of institutional biases in policing in Figure 1b. We find that caste-based biases still appear, as officers believe Brahmins are 4.1pp less likely to be preferred for investigation by the police than Dalits (
$ p=0.026 $
). This suggests that officers acknowledge broader societal norms that position Brahmins at the top of the social hierarchy. And just as we observed that Newars are viewed more like Brahmins than Dalits in our analysis of officers’ personal preferences, there is no statistically meaningful difference in perceived institutional preferences for investigating Brahmins versus Newars. Newars are 1.7pp less likely to be preferred for investigation than Dalits, although this is not a statistically meaningful difference. That said, the negative coefficient suggests directional consistency with officers’ personal preferences. However, despite officers’ personal biases against poor, illiterate offenders, we do not find strong evidence that officers believe these class-based biases exist at the institutional level. Interestingly, individual officers demonstrate certain personal biases that they do not identify at the level of their institution.
What are the interactive effects of CCS backgrounds on investigatory preferences? Figure 2 illuminates the biases more explicitly, rank ordering the likelihood of an individual being investigated for both outcomes as determined by income and caste (Leeper, Hobolt, and Tilley Reference Leeper, Hobolt and Tilley2020). On the left of Figure 2a, we see that a poor Dalit has a 58% probability of being personally preferred for investigation. The right of the figure shows that middle-class Newars are about 14pp less targeted than poor Dalits. Importantly, by being forced to choose between two profiles, the baseline probability of an officer randomly selecting a profile is 50%. Thus, the remaining profile estimates, denoted in gray, are not statistically significant. However, there are meaningful differences between a poor Dalit offender and wealthy Brahmins, Newars, and Dalits. In Figure 2b, we find that officers believe that the police are least likely to investigate middle-class Brahmin offenders. This difference is meaningfully different from poor Dalit offenders.
Officers’ Investigatory Preferences based on Offenders’ Caste-Class Profiles
Note: The probabilities are estimated based on the interaction between income and caste features. The bars indicate 95% confidence intervals and point estimates in black convey statistically significant effects. Point estimates are ordered from highest to lowest. Estimates of panels (a) and (b) of Figure 2 correspond to coefficients inTable A7 in Appendix F of the Supplementary Material.
When exploring heterogeneous effects by respondents’ characteristics, we see that respondent identity affects some investigatory preferences (see Figure A9 and the accompanying discussion in Appendix F of the Supplementary Material). Notably, Brahmin respondents are marginally less likely to prefer investigating wealthy offenders and believe that police will deprioritize investigations of Brahmin offenders. Dalit respondents prefer to investigate and believe that police will prioritize investigating offenders from poor families.Footnote 16 Dalits’ preferences could be due to lack of resources and status to investigate privileged offenders and/or processes of professionalization that pressure Dalits to pledge loyalty to the institution and discriminate against marginalized groups (Blair et al. Reference Blair, Karim, Gilligan and Beardsley2022). When we consider police rank, we find no statistically significant differences on ascriptive attributes between high- and low-ranked officers (see Figure A8 in Appendix F of the Supplementary Material).
Our results are substantively unchanged when we run the same AMCE analyses with sampling weights and/or controls for demographic characteristics (see columns 2–4 and 6–8 of Table A6 in Appendix F of the Supplementary Material). Reassuringly, results do not differ by task order or panel display of profiles (see Figures A6 and A7 in Appendix F of the Supplementary Material). To test the treatment dimension of external validity of our experiment (Egami and Hartman Reference Egami and Hartman2022), we estimate population AMCEs following de la Cuesta, Egami, and Imai (Reference de la Cuesta, Egami and Imai2022), which account for the relative distribution of profiles within the target population as some offender profiles may be more common than others (see Table A9 and Figure A5 in Appendix F of the Supplementary Material). Again, results are substantively similar; however, education and gender are no longer significant predictors.Footnote 17
Asking the police about discrimination raises questions of social desirability bias. Conjoint designs are less susceptible to this bias than traditional survey experiments because presenting multiple attributes simultaneously offers respondents “alternative justifications for sensitive choices” (Hainmueller, Hopkins, and Yamamoto Reference Hainmueller, Hopkins and Yamamoto2014, 19; Horiuchi, Markovich, and Yamamoto Reference Horiuchi, Markovich and Yamamoto2022). Furthermore, these concerns suggest that we would not find evidence of police bias. As such, if social desirability bias is a problem, the effect sizes of the biases we detect are conservative estimates.
CONCLUSION
Absent constraining protocols on what cases to pursue, police officers are more likely to prioritize investigating offenders from CCS communities. Furthermore, officers believe that their institutions are more likely to prioritize the investigation of low-caste offenders; however, they do not view their institutions as having class-based biases, which suggests a potential disconnect between the actions of individual officers and official policy. These findings of the presence of personal investigatory bias and perceived institutional bias add to the body of evidence demonstrating that concerns over systemic biases in policing are warranted. Regardless of whether actual institutional discrimination exists, the disconnect between officers’ stated preferences and perceived preferences of the institution is important: even if they overcome their personal biases, officers may still discriminate if they believe that they are following official policy. Moreover, our finding that personal preferences may not align with perceived views of institutional preferences speaks to the potential importance of recruitment, training, and protocols to reform and abide by institutional rules and norms.
What mechanisms explain this discrimination? We cannot dismiss the possibility that statistical discrimination motivates officers’ preferences, a theory that has come under recent scrutiny for its role in strengthening stereotypes and thereby justifying discriminatory decisions (Spriggs Reference Spriggs2020; Tilcsik Reference Tilcsik2021). However, in the design of our offender profiles, we did not vary their levels of culpability. The language we used stated that all hypothetical offenders had “engaged in” the given crime. Respondents who do not have a “taste” for discrimination should therefore presume that all offenders displayed are equally guilty. Nevertheless, further research is necessary to determine the extent to which taste-based discrimination fully explains the biases we detect, as the preferences we observed could stem from the perception that arrests of certain offenders will more easily result in successful prosecutions or fear of investigating the advantaged.
The journal Perspectives on Politics has called upon political science to take research on police more seriously (Isaac Reference Isaac2015). Soss and Weaver (Reference Soss and Weaver2017) criticize mainstream research in American politics for ignoring the extensive role police play in race-class subjugated communities. Comparative politics has also neglected the police (Crabtree Reference Crabtree2018; Scoggins Reference Scoggins2021; Tanner Reference Tanner2000), although interest has recently increased (see Appendix C of the Supplementary Material). Our study heeds these calls to investigate policing. It speaks to the importance of examining policing bias with respect to which cases are prioritized for investigation, as well as the need to examine institutional discrimination. These findings thus underscore the importance of continued reflection on policing practices globally, with particular attention to the question of how to increase fairness in the criminal justice system.
SUPPLEMENTARY MATERIAL
To view supplementary material for this article, please visit https://doi.org/10.1017/S0003055422001496.
DATA AVAILABILITY STATEMENT
Research documentation and data that support the findings of this study are openly available in the APSR Dataverse at https://doi.org/10.7910/DVN/ESR7YG.
ACKNOWLEDGMENTS
We gratefully acknowledge insightful comments from Miguel de Figueiredo, Gabriel Lenz, and Amy Lerman, as well as from participants of the Annual Meetings of APSA and MPSA; the Woodrow Wilson School Forum at Princeton University; UC Berkeley’s Center on the Politics of Development and the Social Science Matrix. We also thank the researchers of the Latin American Public Opinion Project (LAPOP) at Vanderbilt University for their project and technical support, particularly Nicole Hinton, Daniel Montalvo, Georgina Pizzolitto, and Elizabeth Zechmeister, as well as Tonya Mills for her hard work as our grants manager. We are grateful to Sarah Rich-Zendel, our excellent project lead; Upeksha Tuladhar, our stellar field team member; and Kshitiz Shrestha, who collected our survey data along with his team of enumerators (Deepak Dhungel, Pranita Koirala, Kaji Man Mahatara, Badri Prasad Nepal, and Tara Shrestha) and technical support staff (Jenish Tamrakar). We are indebted to Jenny S. Martinez for supporting this project from its earliest stages. We also thank Maria Barron (USAID/Nepal), Catherine Chen (Humanity United), Lauren Damme and Kevin Hong (Department of Labor), Kristin Farthing and William Mishler (Institute of International Education and USAID), and Jason Squire and Sebastian Zug (Terre des Hommes), who helped make this study possible.
FUNDING STATEMENT
This research was funded by the U.S. Department of Labor, Bureau of International Labor Affairs Grant (#IL-31824-18-75-K). This material does not necessarily reflect the views or policies of our donor and programmatic partners. All errors and opinions are our own.
CONFLICT OF INTEREST
The authors declare no ethical issues or conflicts of interest in this research.
ETHICAL STANDARDS
The authors declare the human subjects research in this article was reviewed and approved by Vanderbilt University. The IRB certificate number and details regarding research ethics are provided in Appendix E of the Supplementary Material. The authors affirm that this article adheres to the APSA’s Principles and Guidance on Human Subject Research.
Open access
Comments
No Comments have been published for this article.