MULTI-STAGE DISCRIMINATION

Most work on discriminationFootnote ² examines isolated stages or “episodic disparities” rather than the reproduction of unequal treatment from one setting to the next (Kurlychek and Johnson Reference Kurlychek and Johnson2019; Kutateladze et al. Reference Kutateladze, Andiloro, Johnson and Spohn2014). The limited theorization and focus on dynamic processes of discrimination suggests that social science is underestimating the true levels of disadvantage that citizens face, including when interacting with linked agencies in a system like criminal justice (Bohren, Hull, and Imas Reference Bohren, Hull and Imas2022; Reskin Reference Reskin2012). Such neglect may yield inaccurate conclusions in scholarship; for instance, if police mishandle or carry out biased investigations, judges have limited evidence to prosecute, and so a single-stage analysis of judicial rulings alone could lead to an imprecise argument that judges are to blame (Lang and Kahn-Lang Spitzer Reference Lang and Kahn-Lang Spitzer2020).

Unlike political science, the disciplines of economics and especially sociology have approached inequality from what Reskin (Reference Reskin2012) calls “a systems perspective,” yet there remains scant dialogue between the fields as to how to conceptualize non-episodic forms of discrimination (Small and Pager Reference Small and Pager2020). Often, terms such as “cumulative discrimination,” “structural discrimination,” “cumulative disadvantage,” and “über discrimination” are interchangeably used to convey similar ideas (Bohren, Hull, and Imas Reference Bohren, Hull and Imas2022; DiPrete and Eirich Reference DiPrete and Eirich2006; Kurlychek and Johnson Reference Kurlychek and Johnson2019; Reskin Reference Reskin2012). Blank (Reference Blank2005, 99) notes that, “cumulative discrimination is the measurement of discriminatory effects over time and across domains,” where one disadvantage feeds into the next. Over-time discrimination may cut across domains (or systems), reinforcing disparities that affect future generations (Lang and Kahn-Lang Spitzer Reference Lang and Kahn-Lang Spitzer2020). Discrimination against people of color in, say, the U.S. housing market can contribute to residential segregation, in turn negatively affecting health or educational opportunities downstream (Korver-Glenn Reference Korver-Glenn2018; Reskin Reference Reskin2012).

I focus on a specific category of “cumulative discrimination” outlined by Blank (Reference Blank2005, 99), that is, “discrimination that arises from multiple interactions within a single social domain over time.” I refer to this as “multi-stage” discrimination because it occurs at sequential nodes within a system or domain (e.g., criminal justice) whereby the process and outcomes associated with a complaint, application, or request are affected at one or more decision-points or “stages.” Despite serving in distinct agencies or sub-systems, administrators that mediate the system’s stages may be inter-connected by rules, routines, or norms. Complaints, applications, or requests travel via stages vertically (through sub-systems, e.g., chain of police custody) or horizontally (across sub-systems, e.g., law enforcement to courthouse). In criminal justice, stages or decision-points might include: (a) police registration, for example, police may turn away or dissuade citizens from case filing; (b) police investigation, for example, officers might delay inquiries or cajole the complainant into withdrawing the report; (c) court trial, for example, judges can stall arbitration or postpone hearings; and (d) court verdict, for example, judges may vary sentences or acquit rather than convict suspects.

Tracing multi-stage discrimination has a defined scope. Broader ideas of cumulative discrimination encapsulate injustices across countless domains which could coincide or overlap at a single time point, or even refer to the experiences of individuals that have not had direct contact with a system. For instance, Black Americans may be victims of discrimination by a U.S. criminal justice system without ever even having officially interacted with the police (Soss and Weaver Reference Soss and Weaver2017). Instead, multi-stage discrimination refers to a specific channel by which inequities propagate across time, conditional on entering a system bounded by rules, routines, or norms. Those that enter this system may have more in common with one another than others that have had indirect or no contact. Predictable stages within the system allow for transitions or serve as leverage points at which administrators have discretion to influence pathways or outcomes.

In a single system, administrators might issue decisions with expectations of how other officials will react (e.g., police may inadequately investigate cases that they know, or perceive, judges will simply dismiss). Feedback loops are particularly salient when administrators are linked by rules, have a joint stake in future decisions, or share resource constraints (Kurlychek and Johnson Reference Kurlychek and Johnson2019). For instance, if an administrator is biased, but is unable to discriminate in the first stage, they may rely on officials downstream to do so. Alternatively, because administrators are constrained by time or resources, they could depend on previous officials in the chain to assist in lightening system load. The mere fact of having multiple stages where human discretion can be applied, absent any coordination by administrators, could lead to greater opportunities for a complaint, application, or request to fail.Footnote ³ Broadly, multi-stage discrimination is only traceable longitudinally rather than at a single stage or time point, and is conceptually distinct from overlapping indirect forms of discrimination that citizens face via disparate domains.

Studying a single system’s stages may shed light on drivers of discrimination. There are numerous mechanisms by which inequalities persist, for example, administrators may have a “taste” for discrimination (Becker Reference Becker2010), they may lack information (statistical discrimination) (Phelps Reference Phelps1972), or the implementation of rules or procedures could “implicitly” or indirectly disadvantage some over others (Bertrand and Duflo Reference Bertrand, Duflo, Banerjee and Duflo2017; Bohren, Hull, and Imas Reference Bohren, Hull and Imas2022). System decomposition could underscore whether certain mechanisms are more likely than others. As an example, suppose biased administrators want to curtail minorities’ cases in early stages but can find no grounds to do so because of the overwhelming evidence. Then, later stages might see a progressively comparable distribution of valid cases. Nevertheless, if minority groups’ requests at a final stage of a system are still more likely to be rejected than the majority—despite having crossed a higher bar of filtration at earlier nodes—we may be better placed in pointing to, say, taste-based versus statistical discrimination on the part of administrators as drivers of injustice (Arrow Reference Arrow2015). In fact, we may even learn about the behaviors of those outside a system too. For example, finding an absence of police discrimination at the same time as citizen anxieties about approaching law enforcement for help could be a puzzle that is clarified by tracing stages; it could be that district attorneys in the middle rungs of a process are discriminatory,Footnote ⁴ which ultimately deters citizens from the first node of case filing with the police. Consequently, exploring within-system discrimination sequentially may reveal more than the sum of its parts at individual stages.

There are other benefits of charting multi-stage discrimination. While experiments causally identify the incidence of discrimination, they do not highlight the process by which disparities compound (Bertrand and Duflo Reference Bertrand, Duflo, Banerjee and Duflo2017),Footnote ⁵ or spotlight where bottlenecks and leakage manifest (Holland Reference Holland2016). Without this fuller picture, any policy intervention applied at one stage might not only fail to reduce inequality, but also have the unintended consequence of exacerbating it. Let us imagine an experiment that reveals statistical discrimination against minorities by judges. Such an empirical finding may drive policymakers to target resources at that node by, say, providing information to administrators so as to minimize future judicial discrimination (Bohren, Imas, and Rosenberg Reference Bohren, Imas and Rosenberg2019; Small and Pager Reference Small and Pager2020). Nevertheless, in tiered systems connected by multiple stages, such one-stop solutions are likely to have limited impact. If judges are provided information, they might update their beliefs or be lenient with related cases in their docket, but patterns of disadvantage will ultimately remain static because of courts’ inability to overturn prior decision-points or even observe cases that they were not sent by law enforcement. Suppose, instead, that it is law enforcement that is discriminatory in terms of preventing cases filed by minority groups from reaching court. Any policy intervention targeted at the node of case filing that reduces or constrains police officer discretion could simply shift discrimination downstream to the judicial stage. Because courts would now be overwhelmed with complaints that they relied on being filtered at earlier nodes, exasperated judges might become more dismissive of minorities’ cases in their docket, thereby intensifying discrimination. In this way, charting a multi-stage process spotlights where anti-discrimination policy interventions should be targeted such that they can have a positive ripple effect through the chain.

Political science scholarship has pointed to inadequacies with the study of (racial) discrimination with administrative (police) data due to selection and incomplete information, for example, the stage at which police stop citizens in the United States (Zhao et al. Reference Zhao, Keele, Small and Joffe2022). As Knox, Lowe, and Mummolo (Reference Knox, Lowe and Mummolo2020) astutely point out, analysts miss information on individuals that police observe but do not investigate, thereby introducing bias in the study of (racial) discrimination by the police. However, decomposing the criminal justice system could be useful vis-à-vis understanding discrimination at later nodes, for example, access to police records may at least shed light on individuals that district attorneys and judges observe but do not prosecute. Still, a multi-stage process induces selection across every node, and studying discrimination within a bounded system does not make it a less challenging statistical task to gain unbiased estimates of the cumulative effect or magnitude of total discrimination (e.g., those with stronger cases might select-in, only those with overwhelming evidence advance, and so on). In fact, if a higher burden of proof is applied to a minority group in an early undocumented stage (e.g., when cases are being reported), the unusually strong complaints reaching subsequent stages could be masking discrimination, especially in administrative records, because officially it might appear that all groups are being treated similarly or even that the disadvantaged group is being treated better.

In this study, I test just one plausible implication of multi-stage discrimination in justice delivery, that is, that disadvantaged groups will see a diminished speed and likelihood of their complaints, applications, or requests successfully crossing the system’s stages. As Kurlychek and Johnson (Reference Kurlychek and Johnson2019) note, research has faced challenges in charting criminal cases across time (and space), even in the United States (Rehavi and Starr Reference Rehavi and Starr2014). I simply follow each administrator decision sequentially in a criminal justice system from the stage of entry (police registration) to exit (judicial verdict) for every complaint. I document whether women face non-episodic unequal outcomes (exclusion) or a disproportionately trying process (burdens) as requests for help transit through the state apparatus of justice provision (Olsen, Kyhse-Andersen, and Moynihan Reference Olsen, Kyhse-Andersen and Moynihan2022).

GENDER AND THE INDIAN CRIMINAL JUSTICE SYSTEM

Research on criminal justice—often restricted to the United States, and focused on the police or judiciary—shows that Black Americans are discriminated against in terms of bail, sentencing, and incarceration (Abrams, Bertrand, and Mullainathan Reference Abrams, Bertrand and Mullainathan2012; Alesina and La Ferrara Reference Alesina and La Ferrara2014; Arnold, Dobbie, and Yang Reference Arnold, Dobbie and Yang2018). Aside from related scholarship on ethnicity (Shayo and Zussman Reference Shayo and Zussman2011), there are few analyses of gender-based disparities. I address this lacuna by focusing on India; here, women may have limited support such as access to lawyers (Roychowdhury Reference Roychowdhury2021), and investigating VAW might be perceived as a strain on police resources. Culturally, officials might construe punitive justice for women’s complaints, including dowry,Footnote ⁶ as a threat to societal norms or marriage. Further, because gender disparities are pronounced in areas like education or labor force participation, we may see imbalances between men and women vis-à-vis accessing justice even absent any malign intent by administrators, for example, women could be fearful about retribution and humiliation or lack autonomy from the household to travel to courthouses (Chhibber Reference Chhibber2002).

The Hindi-speaking belt of north and central India is a crucial site to study inequality (Jayachandran Reference Jayachandran2017); by global standards, this region, and especially Haryana state, retains among the most skewed sex-ratios in favor of men (834 girls to 1,000 boys in 2011 [Chhibber, Jensenius, and Ostermann Reference Chhibber, Jensenius and Ostermann2021]) as a function of sex-selective abortion and female infanticide. Haryana may thus present not only an upper bound of gender-based disparities in accessing justice, but also illustrate how inequalities reproduce in the formal system of grievance redressal.

When seeking help from the state to address a wrong, crime registration is a primary step. It occurs at police stations run by a station officer, who is supported by staff (e.g., sub- and assistant sub-inspectors). Police are supposed to file all complaints whether they believe them to be truthful or not, but in practice have leeway as to which are formally registered. When filed, the station officer assigns a case to a deputy, and, depending on crime-type, investigations have to be completed within a window (e.g., 60 days). If not canceled, or withdrawn, the case is sent to court. Every station is located within a jurisdiction of a district court; police reports, and any evidence collected during investigation, are assigned to a judge.Footnote ⁷ In a major study that introduces a fine-grained dataset of Indian court decisions, Ash et al. (Reference Ash, Asher, Bhowmick, Chen, Devi, Goessmann and Novosad2021) find a null effect of (judicial) bias with regard to in-group defendants. Here, I focus not those who have been accused of crime per se, but on those who initiated the legal action in the previous stage: the plaintiffs.

Figure 1 presents a stylized illustration. Level A represents the abstract concept of all crime. Level B signifies those who came forward to report, for example, at a station or help-desk (Sukhtankar, Kruks-Wisner, and Mangla Reference Sukhtankar, Kruks-Wisner and Mangla2022). Level 1 is smaller than “B” because not all reported cases are registered. This study focuses on those cases that were filed with the state. Within level 1, there are two sub-categories: women’s complaints and VAW.Footnote ⁸ This is illustrated in a Venn diagram because not all VAW is reported by women.Footnote ⁹ Cases in level 2 represent those that, after registration, survive cancelation. The remaining cases, once investigated, enter the judiciary in level 3. There, unless stalled or dismissed, cases result in a verdict following trial that (dis)favors the complainant in the original filing from level 1 (who eventually became the plaintiff in court).

Figure 1.

Levels of Accessing Formal Justice in India

Note: Light and dark blue represent police jurisdiction; yellow represents the judiciary. The analyses in this study cover all levels from 1 to 3, and the corresponding in-between stages.

Judges arguably have greater discretion when handling cases than the police. For officers, there are explicit rules that mandate registration of “cognizable” or serious crimes,Footnote ¹⁰ some introduced after a gruesome 2012 gang-rape of a Delhi college student. Police are required to register all VAW complaints—including acid attacks, criminal force, trafficking, and rape—with the threat of 1-year jail time or fine for the officer, and rape investigations are to be completed within 2 months of filing.Footnote ¹¹ Aside from being pressured “from above” via such guidelines and orders, the police are also constrained “from below” where, for example, feminist groups and NGOs can assist victims in filing cases, especially VAW (Htun and Weldon Reference Htun and Weldon2012). Judges are disassociated from such pressuresFootnote ¹² or from juries, which were formally abolished in 1973.Footnote ¹³

During registration, police officers stamp Penal Codes to cases to signal the probable laws violated. VAW Penal Codes (and related “acts”) include Section 326-A (acid throwing), Section 498-A (dowry harassment),Footnote ¹⁴ Section 376 (rape),Footnote ¹⁵ and others.Footnote ¹⁶ Politicians have argued that women exaggerate when filing such cases, even noting, “Many families are destroyed or ruined under such [gendered] provisions, and the legal proceedings go on for years. Men’s rights organizations are working to raise awareness … in opposition to women … men should be arrested after proper inquiries rather than on the basis of the woman’s complaint” (Verma Reference Verma2017).

These sentiments are not restricted to politicians. (All-male) benches of the Supreme Court have ruled that domestic violence provisions are, “a license for unscrupulous persons to wreck personal vendetta or unleash harassment [against men],” and a form of “legal terrorism [by women].”Footnote ¹⁷ The court has noted, “… complaints under Section 498-A are filed in the heat of the moment over trivial issues without proper deliberations. The learned members of the Bar have enormous social responsibility and obligation to ensure that the social fiber of family life is not ruined or demolished,”Footnote ¹⁸ and that women should be deterred from filing cases to, “satisfy the ego and anger of the complainant.”Footnote ¹⁹ These statements imply that cases of VAW are (a) frivolous, (b) reported without delay, (c) submitted by those with an agenda, or (d) best resolved through reconciliation (Basu Reference Basu2012; Jassal Reference Jassal2021). I scrutinize these assumptions using two sources of data.

THE FIRST-INFORMATION-REPORT DATASET AND JUDICIAL RECORDS

In a push for transparency, India made police reports, called First-Information-Reports (FIRs), accessible to citizens. Over several years, I collected and parsed millions of these records; this study utilizes 418,190 police files in Haryana from January 2015 to November 2018.Footnote ²⁰ I focus on this state for which I translated reports into English, and was able to collaborate with the local police to gather information about officers and otherwise withheld cases.Footnote ²¹ Aside from details about victims and suspects, FIRs contain first-person descriptions of crime, less affected by social desirability.Footnote ²² I then merged FIRs with judicial records. India has made (semi-)public the universe of these files on a platform called E-Courts, similar to a domain established by China (Liebman et al. Reference Liebman, Roberts, Stern and Wang2020). It contains particulars about the date of filing/first appearance in court for a police report, judges assigned, verdict (if any), and other details. With assistance from the Development Data LabFootnote ²³—that compiled and released 77 million records from 2010—I connected the two databases via particulars of the station, complainant name,Footnote ²⁴ and other identifiers. Out of 418,190 crime reports, I merged precisely 251,804 or 60.2% to court files, a figure that appears to accurately reflect those police files that were ultimately sent to the judiciary.Footnote ²⁵

Research Design

I chart the process and outcomes associated with each level of accessing formal justice (Figure 2). In level 1 (registration), I examine the duration of time to file a crime report. Crime records include when the complaint was filed, as well as the dates that a complainant told an officer that the crime began or ended. Registration duration reflects the difference (in days) between police filing and incident. For outcome, I examine the likelihood of a registered case being sent to court; non-merged cases are categorized as canceled, indicating that the police files were not present in the judiciary. For levels 2–3 (investigation and preliminary hearing), I create two measures. Investigation duration—the difference (in days) between police filing and the case’s first appearance in court—signifies how long the police inquiries took. Dismissal refers to whether the case was ejected by a judge at the initial hearing. At the final stage (level 3), I create a numeric variable corresponding to the number of days from the preliminary hearing until the most recent one on file (duration in court). I focus on two indicators for judicial review, that is, whether the verdict issued by a judge resulted in a suspect’s conviction and acquittal.

Figure 2.

Process and Outcome Measures of Accessing Justice in India

Broadly, I am interested not only in the primary question as to whether women are disadvantaged while accessing justice compared to men (including for generic or non-VAW cases), but also whether male complainants are less likely to face burdens and exclusion when registering cases of VAW on behalf of women. In the equation below, Y_i is a binary or numeric outcome for crime report i filed in police station s at time t:

(1) $$ {Y}_i={\displaystyle \begin{array}{l}\alpha +{\beta}_1Femal{e}_i+{\beta}_2VA{W}_i+{\beta}_3\Big(Female\\ {}{\cdot VAW\Big)}_i+{X}_i\gamma +{\zeta}_s+{\upsilon}_t+{\epsilon}_{i.}\end{array}} $$

$ Femal{e}_i $ indicates whether the case was filed by a woman, and $ VA{W}_i $ is a binary variable for any Penal Code classified a “crime against women” being affixed to the report. Since VAW may be registered not only by women but also by male family/friends, the interaction term allows for differences between male and female complainants for VAW and non-VAW crime. $ {X}_i $ represents covariates, for example, distance in kilometers that the crime took place from the station, rank of the officer that investigated the case, and, ultimately, rank of the judge that heard it in court. I also include fixed effects for the station that the case was registered in ( $ {\zeta}_s $ ) and the month–year of police filing ( $ {\upsilon}_t $ ). In the Supplementary Material, when excluding $ VA{W}_i $ , I include primary Penal Code fixed effects.Footnote ²⁶ Standard errors are clustered at the district-level.

I also estimate structural topic models (STMs) that, in a regression-type framework, can predict whether cases devoted to a topic (e.g., rape) are functions of covariates (Roberts, Stewart, and Airoldi Reference Roberts, Stewart and Airoldi2016; Roberts, Stewart, and Tingley Reference Roberts, Stewart and Tingley2019; Roberts et al. Reference Roberts, Stewart, Tingley, Lucas, Leder-Luis, Gadarian and Albertson2014).Footnote ²⁷ The method de-emphasizes Penal Codes by textually disaggregating what citizens told the police happened to them using statistical associations between words to separate, say, domestic violence from domestic violence that also happened to involve attempted murder (e.g., an inexperienced officer could have incorrectly classified a case or have forgotten to append a relevant Penal Code). I compiled and parsed text from each police report into a machine-readable format, and translated the 418,190 cases (two hundred million words or $ \approx $ 450,000 A4-size single pages) using Google Translate.Footnote ²⁸

When comparing men and women’s attempts at accessing justice, there may be concerns about omitted variable bias and distinct underlying case distributions (Arnold, Dobbie, and Yang Reference Arnold, Dobbie and Yang2018), for example, even within categories of crime (theft), women may report distinct sub-types (bag-snatching) compared to men (motorcycle robbery). Consequently, I utilize a third method: topical inverse regression matching (TIRM), introduced by Roberts, Stewart, and Nielsen (Reference Roberts, Stewart and Nielsen2020), which allows for the matching of complaints based on text, thereby adjusting for confounding. To implement TIRM, I estimate a STM with an indicator for a woman’s crime report as a content covariate. This estimates the relationship between having a female complainant and words in the corpus, as well as how crime reports registered by women discuss topics differently (Roberts, Stewart, and Airoldi Reference Roberts, Stewart and Airoldi2016). Following the procedure outlined in Roberts, Stewart, and Nielsen (Reference Roberts, Stewart and Nielsen2020), I create indicators for whether the police report ultimately resulted in a suspect’s conviction and acquittal in court. Put differently, I test if final stage outcomes are distinct for men and women who brought forward topically similar complaints at the very first stage. In this setting, the method likely serves as the closest one may attain in terms of having a valid counterfactual, short of randomizing citizens to be victims of crime or report at police stations.

DESCRIPTIVE STATISTICS

Women registered 38,828 or 9% of all FIRs. Supplementary Figure A2 displays the top Penal Codes in women’s cases, and Supplementary Figure A1 highlights those for men. For male complainants, the top substantiveFootnote ²⁹ Penal Codes relate to theft, rash driving, burglary, and public intoxication/bootlegging. For women, Section 498-A or domestic violence/dowry-related abuse perpetrated by a spouse (or in-laws) was present in 15% of their registrations.Footnote ³⁰ Other common VAW Penal Codes include abduction (e.g., kidnapping a woman “to compel her into marriage”),Footnote ³¹ “obscene acts/songs,”Footnote ³² “criminal force against a woman,”Footnote ³³ rape, “insulting the modesty of a woman,”Footnote ³⁴ stalking, “intent to disrobe,” sexual harassment, and “unnatural” (anal) sex.Footnote ³⁵

Table 1, which depicts the first quantification of Indian cases, presents summary statistics. Distance reveals that crime takes place, on average, 5.5 kilometers from a station. When identified by the complainant, cases are likely to have two suspects. Crimes registered by women, and VAW, are more likely to have a female suspect (female suspects). (Dowry may involve a mother-in-law.) While officers do not always record victim ages, non-missing data suggest that complainants are, on average, in their 30s. VAW is likely to have more Penal Codes (no. of sections). Strikingly, victims of VAW wait longer at the station in anticipation of registration (9 vs. 7 hours).Footnote ³⁶ Women’s cases are infrequently assigned to junior officers, for example, constables.

Table 1.

Descriptive Statistics on Select Variables: The First-Information-Report (FIR) Dataset

Note: Descriptive statistics for variables in the FIR dataset, split by female/other complainants, as well as VAW/non-VAW crime. The term ‘Other’ is used because a small fraction of cases may be brought forward by organizations or institutions rather than individuals. VAW crime may be brought forward by male or female complainants. Variables prefixed with ‘R:’ represent investigator ranks.

The median days between crime and registration (registration duration) is 1, with a mean of 28. Women’s cases, and VAW, have means of 69 and 113, respectively. A complainant may have visited a station to register a case but asked to drop it, or be forced to return at a later date.Footnote ³⁷ Prima facie, this challenges the assumption that women’s cases are filed, “in the heat of the moment.”Footnote ³⁸ No record shows that women’s cases are 2 percentage points more likely to be canceled at the police-level.Footnote ³⁹ Table 2 highlights variables post-merging with court files. Investigation duration shows that a police investigation takes 128 days, on average. Cases spend about 336 days in the judiciary (duration in court); yet, women’s cases, and VAW, spend even longer awaiting a decision. The variable duration in court is noteworthy when juxtaposed with the comparable number of court meetings for male and female complainants (no. of hearings). While most cases are assigned to Judicial Magistrate 1st Class, women’s cases, and VAW, are more likely to be assigned to senior judges, for example, Additional District Sessions Judge.Footnote ⁴⁰

Table 2.

Descriptive Statistics: The First-Information-Report (FIR) Dataset Merged with Court Records

Note: Descriptives statistics for select variables in merged dataset of crime and judicial records, split by female/other complainants, as well as VAW/non-VAW crime. The term ‘Other’ is used because a small fraction of cases may be brought forward by organizations or institutions rather than individuals. VAW crime may be brought forward by male or female complainants. Variables prefixed with ‘R:’ represent judge ranks.

Figure 3 illustrates judicial outcomes, which fall into roughly seven categories in the E-Courts database. Acquitted refers to whether the suspect is absolved; allowed denotes if the case is admitted but a trial has not been set; convicted represents suspect conviction, whereas dismissed underscores a judge’s ejection of the case, typically at a bail hearing.Footnote ⁴¹ The basic cross-tabulations in Figure 3 show that—whether a function of all FIRs (panels a and b) or simply those in the court docket (panels c and d)—women’s complaints (and VAW) are more likely to be on-going (stalled), dismissed, or result in a suspect’s acquittal, and less likely to see a suspect sent to prison. For women’s cases, only 2.9% of the implicated suspects are convicted, unlike 10.8% for men’s cases. For cases that ultimately make their way to court, the gap between female and male complainants’ cases is even wider (5% and 17.9%, respectively).

Figure 3.

Crime Report Statuses [Split by Complainant Gender and Crime Type]

Note: Judicial outcomes for cases (% on Y-axis). Panels a and b reflect outcomes conditional on police registration, that is, including “no record” cases or police files that did not make their way to court. Panel a is separated by female (N=38,828) and male/other complainants (N=379,362). Panel b reflects VAW (N=20,869) and non-VAW crime (N=397,321). Panels c and d reflect outcomes as a function of cases just in the court docket. Panel c is separated by female (N=22,648) and male/other complainants (N=229,156), and Panel d by VAW (N=14,134) and non-VAW crime (N=237,670). 95% confidence intervals included.

OLS RESULTS

While there appear to be gender gaps along a variety of measures, women are also more likely to register cases involving VAW which—for political, economic, and cultural reasons—could be treated specially by the criminal justice system. And so, I examine how women fare in each stage of justice delivery, taking into account this confounding factor.

Columns 1 and 2 of Table 3 show that women’s cases have, on average, a lag of over a month longer than men between incident and registration, suggesting significant delays between crime occurrence and when the state takes cognizance of the case. Columns 3–6 show that cases of VAW account for a substantial portion of this gap. While this may be reflective of women’s hesitancy in coming forward to complain of such crime, it is important to note that at the stage prior to cases formally entering the books, law enforcement has discretion in asking citizens to return at a later date or forwarding complainants of VAW to counseling and mediation centers. Still, controlling for VAW, columns 5 and 6 reveal significant discrepancies between men and women for generic cases; for non-VAW, the number of days from when a crime occurs and police officers initiating the case for investigation is roughly a week longer for women than men.

Table 3.

Police Process and Outcome for Female Complainants: Level 1

Note: Female indicates whether the complainant in the police report was a woman. Controls include a numeric variable for distance of crime from station and investigator rank. Standard errors clustered by district. For full model, see Section 7 of the Supplementary Material. *p < 0.1; **p < 0.05; ***p < 0.01.

In terms of outcome in level 1, columns 7 and 8 of Table 3 reveal that women’s cases are significantly less likely than men’s to be sent to court. However, conditional on registration, cases of VAW are likely to transition to the next wing compared to non-VAW (columns 9 and 10), especially when a woman is the complainant (columns 11 and 12). Cases of VAW are approximately 7–8 points more likely to be sent to the judiciary than non-VAW crime. To reiterate, police officers are bound by rules to ensure (registered) cases of VAW transition or are investigated quickly. The gap between men and women in terms of a case being canceled by law enforcement (after registration) is noteworthy for non-VAW, that is, those complaints for which officers have discretion in influencing outcomes. This dynamic is reflected in terms of police investigations too. Specifically, columns 3 and 4 of Table 4 show that cases of VAW are investigated and sent to court, on average, sooner than non-VAW crime. And, while women’s cases are investigated slower by police officers than men’s complaints (column 2 of Table 4), this is especially true when the cases are restricted to non-VAW where women’s complaints have investigative delays of $ \approx $ 19 days (columns 5 and 6).

Figure 4 presents marginal effects in a descriptive plot. Panel a shows that cases of VAW (registered by women) have the longest lag between incident and filing. However, having crossed the node of filing, VAW is generally allowed to pass through level 1 (panels b and c).

Figure 4.

Marginal Effects for Male and Female Complainants across Stages (for VAW and Non-VAW)

Note: Marginal effects based on regressions in columns 6 or 12 in Tables 3–6.

Table 4.

Police/Judicial Process and Outcome for Female Complainants: Level 2

Note: Female indicates whether the complainant in the police report was a woman. Controls include a numeric variable for distance of crime from station, investigator rank, and judge rank. Standard errors clustered by district. For full model, see Section 7 of the Supplementary Material. *p < 0.1; **p < 0.05; ***p < 0.01.

Nonetheless, discrimination appears more consistent—whether in contexts of VAW or not (columns 7–12 of Table 4 and Figure 4d)—when cases selected to leave the police jurisdiction enter the court for the first time. Specifically, even though women’s non-VAW cases are between 1 and 3 points more likely to be dismissed at a preliminary court hearing than cases brought by men, a gap persists for VAW. Figure 4d–f suggests that not only are women discriminated against across crime type, but also male complainants who register cases on behalf of female friends or relatives are less likely to face burdens or exclusion than if a woman was listed as the primary complainant. Columns 1 and 2 in Table 5 reveal that women’s cases spend longer in the judiciary by over a month. Graphically, Figure 4e shows that cases involving VAW registered by women spend significantly long stalled. To probe whether punitive justice was meted out or that the individual who sought help in the original report received a favorable ruling, I pay attention to the suspect’s conviction or acquittal who allegedly wronged the complainant. Columns 1, 2, 5, and 6 of Table 6 demonstrate that cases brought forward by women are significantly more likely to yield a suspect’s acquittal. Women’s cases are associated with more than a 10-point reduction in convictions (columns 7, 8, 11, and 12). Figure 4f shows conviction for suspects that male complainants accuse of VAW (e.g., for female family/friends) decline from a non-VAW base, but not to the same level as when women file cases themselves. Indeed, women complainants seeking justice from the state have a lower chance of a suspect that wronged them being sent to prison for either type of complaint, VAW or not.

Table 5.

Judicial Process for Female Complainants: Level 3

Note: Female indicates whether the complainant in the police report (or plaintiff in court) was a woman. Controls include a numeric variable for distance of crime from station, investigator rank, and judge rank. Standard errors clustered by district. For full model, see Section 7 of the Supplementary Material. *p < 0.1; **p < 0.05; ***p < 0.01.

Table 6.

Judicial Outcomes for Female Complainants: Level 3

Note: Female indicates whether the complainant in the police report (or plaintiff in court) was a woman. The dependent variable refers to whether the suspect that was implicated in the crime was acquitted or convicted, respectively. Controls include a numeric variable for distance of crime from station, investigator rank, and judge rank. Standard errors clustered by district. For full model, see Section 7 of the Supplementary Material. *p < 0.1; **p < 0.05; ***p < 0.01.

While this paper does not delve into drivers of discrimination, the findings are suggestive. On the one hand, administrators may be acting rationally but with imperfect information. Administrators could hold inaccurate beliefs about women’s tendency to exaggerate. Or, female complainants may be statistically less likely to afford lawyers or cope with in-person follow-ups necessary for trial, thereby precluding judges from making informed decisions. (In India, complainants have to be present in court at least twice, e.g., once during a magistrate’s “cognizance” of a case, and again during cross-examination.) Describing crime in open court can be difficult, especially for victims of VAW. Further, traveling long distances repeatedly to a district court for multiple hearings may pose distinct challenges than simply filing a one-time police report at a neighborhood police station. Low levels of development can result in misgovernance without administrators behaving with repressive intent or harboring animus (Slough and Fariss Reference Slough and Fariss2021).

Then again, administrators may have a taste for discrimination or, say, eager to “protect” victims from the complex and public process of accessing formal justice (Bindler and Hjalmarsson Reference Bindler and Hjalmarsson2020). The empirical findings suggest that statistical discrimination is unlikely to be the only mechanism, since it is plausible that the gaps would likely decrease across later stages as a function of the initial filtration. For example, Stolzenberg, D’Alessio, and Eitle (Reference Stolzenberg, D’Alessio and Eitle2013) note that racial discrimination in the U.S. justice system is less likely to manifest at final decision-points because initial processing decisions ultimately decrease variation, that is, cases in later stages become progressively similar vis-à-vis the validity of evidence. Here, gaps amplify, persisting in the final node of a judge’s verdict, a point at which plaintiffs do not have to be physically present or travel long distances. Relatedly, if police were dismissing complaints without animus by factoring in expected judicial rulings for what they perceived would be statistically hard-to-prove cases in court, then officers might presumably apply discretion to VAW.Footnote ⁴² Instead, gaps at the police-level are stark for non-VAW crime, that is, cases for which officers have no prima facie reason to anticipate negative consequences in court to warrant differential treatment. And so, it is likely that taste-based discrimination operates in conjunction with other drivers of inequality.

In Supplementary Table A2, I re-run the analyses controlling for over a thousand Penal Codes/acts. I also analyze the outcomes for levels 2 and 3 as a function of all registrations (as opposed to just the court docket) to account for the selection that occurs when only 60% of police files are sent to the judiciary (Supplementary Tables A7 and A8 and Supplementary Figure A27). The coefficient on female remains significant in almost every model. A positive implication of the findings is that, conditional on registration, cases of VAW have better outcomes in in the realm of the police (e.g., plausibly as a function of the rules and guidelines that check officer behavior), but there is also suggestive evidence that, because of the multi-stage process, discrimination is delayed rather than mitigated. Discrimination does not appear to be restricted to one agency, but is iterative, and present in the mid- to late-stages of justice delivery, “the last mile” at which complainants have not only spent time and resources to reach those stages but also decision-points at which there are few constraints on administrators from either “above” or “below.”Footnote ⁴³

Structural Topic Modeling (STM)

There remain at least two methodological concerns with the analyses thus far. First, case categorizations have relied on Penal Codes. The assignment of codes is a strategic decision by officers who may, consciously or unconsciously, underweight a case’s seriousness by deciding which and how many to apply.Footnote ⁴⁴ Second, even if one were to accept that there is an imbalance between men and women, perhaps women are more likely to register baseless cases, which the justice system happens to be efficiently weeding out. To explore the validity of this assumption, I utilize the unfiltered first-person testimony that citizens provide to law enforcement prior to case investigation, thereby setting aside the administrators’ classifications or any official coding schema.

I apply STMs to reports. Such modeling can estimate relationships between meta-data and topics from the corpus and facilitate hypothesis testing (Roberts, Stewart, and Tingley Reference Roberts, Stewart and Tingley2019). Are there topics in the victims’ testimonies—including inside women’s complaints—that yield low convictions for suspects? To reiterate, the goals of STM are three-fold: (1) give voice to victims by utilizing their own words, (2) highlight the severity of claims, especially VAW, and (3) coarsen high-dimensional data to allow for text matching techniques.

I begin by summarizing crime in north India. In Supplementary Figure A29, there are crimes that are likely to be familiar to most readers including public intoxication/bootlegging or “alcohol” (Topic 19), “burglary” (Topic 16), “auto theft” (Topic 22/23), and “kidnapping” (Topic 27). However, there are South Asia-specific cases such as Topic 5 “cattle,” which represents the smuggling of cows or their slaughter, and Topic 6 “resource mafia” that signifies the sand or mining cartel (Asher and Novosad Reference Asher and Novosad2023). Figure 5 utilizes an indicator for suspect conviction as a predictor, and shows correlations (when topics likely co-occur). Across the full corpus, there are topics that appear more likely to result in suspect conviction, for example, fraudulent currency, gambling, and alcohol-related cases (Topics 2, 19, 28), compared to others like auto-theft, missing persons, and kidnapping (Topics 22, 7, 27), for which suspects infrequently go to prison or are even found. In Figure 5b, the machine estimated correlations where, for instance, Topics 10 and 13 (“accident” and “injury”) are unsurprisingly related, as are “cattle” and “minorities,” suggesting that members of the Muslim community are disproportionately victimized (or, more precisely, invoked in a police report) for alleged bovine-related offenses. More research is needed to explain the heterogeneity, and why, for instance, topics related to “arms” (Topic 18) and organized crime (Topic 6) have better indicators on suspect acquittal (Supplementary Figure A31).

Figure 5.

Suspect Conviction and Correlation of Topics Associated with Full Crime Corpus

Note: Panel a: Coefficients and standard errors for a structural topic model of all police complaints filed in Haryana with suspect conviction/non-conviction in court as the predictor. Right of the dashed vertical line represents positive coefficients. The stemmed words making up the topics appear in Supplementary Figure A29. Panel b: Figure depicts the network of correlated topics. Colors indicate the magnitude of the coefficient; red underscores positive coefficients and blue negative for the suspect conviction indicator. The gray widths of the edges are proportional to the strength of correlation between topics.

More importantly, the exercise sheds light on gradations of abuse that women face.Footnote ⁴⁵ Figures 6 and 7 present the highest probability as well as FREX (frequent and exclusive) words for women’s complaints and VAW, respectively. A top topic that emerges involves “fighting” (Topic 14), usually domestic violence. The word clouds in Supplementary Figures A37–A39 underscore terms such as: wife, hospital, kill, beaten, domest, husband, hurt, and blunt. Figure 7 breaks down VAW.Footnote ⁴⁶ Topics range from the blackmail of women with compromising photographs/videos (Topic 18) to “trafficking” or being sold into prostitution (Topic 12). While certain topics such as sexual assault (by a non-spouse) or child abuse have better outcomes (vis-à-vis suspect conviction) (Figure 9), a theme that emerges is the prioritization of sons over daughters. Specifically, in Figure 7, Topic 7 refers to abandoning or killing infants (“killing the girl child”), Topic 14 refers to (illegal) sex selective diagnostic technologies, and Topic 5 includes unlicensed doctors performing abortions. As highlighted in the word clouds of the Supplementary Material (Supplementary Figures A42 and A44), common words in these categories include: children, child, medic, drug, abort, kill, patient, ultrasound, and pregnant.

Figure 6.

Top Topics (Female Complainants, N=38,828)

Note: Panel a: Top topics associated with women’s complaints and highest probability words in the topic. Panel b: FREX words (frequent and exclusive) or distinguishing words of the topics.

Figure 7.

Top Topics (VAW Crime, N=20,869)

Note: Panel a: Top topics associated with cases of violence against women (VAW) and highest probability words in the topic. Panel b: FREX words (frequent and exclusive) or distinguishing words of the topics.

STM depicts how dowry underlies various forms of VAW. In both Figures 8 and 9, we see clusters of topics related to violence involving dowry-based harassment and domestic abuse. As the blue shading suggests, almost all these forms of crime are unlikely to see suspect conviction, for example, Topics 5, 6, 9, 13, and 23 in Figure 9.Footnote ⁴⁷ Yet, based on the words that complainants use to describe this abuse, the cases do not appear to be frivolous. Supplementary Figures A37–A39 show that common words include: dowry, tortur, parent, cash, daughter, greed, kill, demand, cruelti, in-law, and assault. “Mother-in-law” appears repeatedly, indicating that abuse often involves the extended family as opposed to just an intimate partner. The machine can separate dowry relating to mental or physical torment in Figure 7 (Topics 1 and 2) from others involving, for instance, harassment in conjunction with spousal rape (Topic 6). This is highlighted in the FREX words in Figure 7b that accentuate terms such as unnatur (i.e., “unnatural” or anal) and sexual. Footnote ⁴⁸ Topic 16 represents instances in which complainants explain that they tried to register a dowry or domestic violence case before but were asked to reconcile or participate in counseling with a spouse instead. This is seen in the FREX words mediate or counsel in panel b (Jassal and Barnhardt Reference Jassal and Barnhardt2023). Topics 19 and 20 refer to abusers either deserting their wives or absconding, for example, possibly to extract dowry from another. The only type of dowry-related topic that is associated with higher levels of suspect conviction is Topic 8, that is, when harassment has culminated in suicide or the death of a victim (equivalent to murder); this is evoked in the FREX words of Figure 7, for example, dowri, die, marri, death, hang, poison. The severity of cases which crossed the first stage of registration involving physical, mental, and emotional abuse that spouses (and in-laws) perpetrate, often to extort resources from victims’ natal homes, go against assertions that such complaints are family disputes unworthy of formal punishment by the state.Footnote ⁴⁹

Figure 8.

Suspect Conviction and Correlation of Topics Associated with Women’s Cases

Note: Panel a: Coefficients and standard errors for a structural topic model of police complaints filed by women with suspect conviction/non-conviction in court as the predictor. Right of the dashed vertical line represents positive coefficients. The stemmed words making up the topics appear in Figure 6. Panel b: Figure depicts the network of correlated topics. Colors indicate the magnitude of the coefficient; red underscores positive coefficients and blue negative for the suspect conviction indicator. The gray widths of the edges are proportional to the strength of correlation between topics. Gradations of VAW appear highly correlated with each other (top of panel b), while driving accidents/hit-and-runs have limited connections to other types of crime.

Figure 9.

Suspect Conviction and Correlation of Topics Associated with Violence against Women (VAW) Crime

Note: Panel a: Coefficients and standard errors for a structural topic model of police complaints involving VAW (filed by women or male friends/family of victim) with suspect conviction/non-conviction in court as the predictor. Right of the dashed vertical line represents positive coefficients. The stemmed words making up the topics appear in Figure 7. Panel b: Figure depicts the network of correlated topics. Colors indicate the magnitude of the coefficient; red underscores positive coefficients and blue negative for the suspect conviction indicator. The gray widths of the edges are proportional to the strength of correlation between topics. Gradations of dowry/domestic violence cases appear highly correlated with each other. Sexual assault by a non-spouse is (relatively) more likely to lead to suspect conviction (Topic 17) than marital rape (Topic 6).

Topical Inverse Regression Matching (TIRM)

STM underscores yet another challenge with interpreting the regression analyses and selection, for example, men and women’s cases could be distinct in a way that controlling for Penal Codes or other variables across the system cannot account for. For women, a common form of theft is “chain-snatching” (Topic 15 in Figure 6), as opposed to vehicle robbery for men (Topic 22 in Supplementary Figure A29). While an officer would have simply classified both as “theft” (e.g., Section 379), one could assume that the criminal justice administration does not discriminate against women per se but merely takes stolen vehicles more seriously than burgled jewelry. One way to tackle this problem is via text matching utilizing all original police reports; then, after qualitatively ensuring that the technique was successful, compare final outcomes (Grimmer and Stewart Reference Grimmer and Stewart2013). Of course, even thinking about gender conceptually as a “treatment” is difficult; it is a non-manipulable “bundled” category encapsulating numerous factors (Holland Reference Holland1986; Knox, Lowe, and Mummolo Reference Knox, Lowe and Mummolo2020). Nevertheless, I hold features of the testimony provided to law enforcement constant but for the complainant’s gender in a novel but imperfect attempt at underscoring a plausible causal link between complainant identity and justice.

I utilize registered complaints for topical inverse regression matching. Figure 10 is the first balance-test. The gray bars—which highlight the difference between female minus male complainants’ topics in the unmatched data—reveal differences by gender. Women are more likely to register dowry violence (Topic 24), whereas men cases of bootlegging and alcohol (Topic 4). Figure 10 shows that TIRM (black dots) is highly successful in minimizing these differences.

Figure 10.

Balance Check I

Note: Gray bars indicate cases associated with male and female complainants. For instance, women complainants are likely to bring forward dowry cases (e.g., Topic 24), while male complainants are involved in or file alcohol and bootlegging (Topic 4). TIRM tries to achieve balance on estimated topics (black dots).

As a second balance test, I randomly select and present 12 matched testimonies in Table 7. This is a hard test for balance because the machine matched cases without any reference to Penal Codes; and still, post-TIRM, we see similarities in codes based on content. In fact, the machine is more successful at categorizations than police officers.Footnote ⁵⁰ In rows 1, 2, 5, and 6 of Table 7, we see non-VAW cases registered by either a male or female complainant (identifying information redacted). Row 1 depicts scooter theft, and row 2 a hit-and-run. In row 2, we see matched cases where a crash occurred. Still, despite being similar, there remain dissimilarities that the machine cannot (and, in fact, should not) match on. In row 6, for both men and women, the cases involve confidence-tricksters, but the type of con is distinct. Relatedly, the woman’s case in the dowry murder involves a killing, but in the matched column, a woman and her child have been found deceased. In the abduction cases, the complainants belong to distinct religious communities (Hindu and Muslim, respectively).

The language in the documents is rich, and allows for a brief interpretative exercise. In row 3 of Table 7, we see (relatively less violent) dowry cases wherein victims have been extorted, and in the left column of the table, beaten. Consider the way in which class is foregrounded. In the right column, the father—who is registering a case on behalf of his child—notes that his daughter is well-educated. He also underscores his social humiliation. The complainant in the left column of Table 7 is filing a case against a lawyer and judge, which suggests not only that the perpetrators have influence, but also that they are well-educated; yet, the suspects allegedly believe that they are owed luxury vehicles in view of their “status.” Similarly, in row 4, the complainant in the left column notes that the in-laws (in an arranged marriage) were given gifts in accordance “with their status.”Footnote ⁵¹ These dynamics suggest that class plays a role in mediating interactions; even with officers, citizens signal their background, potentially to be taken seriously. Often, citizens do not have cultural capital, and are forced to plead for help (as in the right column for the “abduction” row where the complainant stresses his poverty). A puzzle arises as to how justice would vary across these contexts: would the system provide re-distributive justice (financial compensation), especially for losses in the dowry or cheating cases? Would those with political connections or social capital be more successful in seeing their requests cross the system’s stages?

In the left column of row 4 in Table 7, the perpetrators previously went to prison. This raises concerns about the failure to deter perpetrators such that a woman was allegedly murdered despite the suspects’ prior incarceration. The reports shed light on criminal impunity, where individuals may be abducted in broad daylight, or killed in defiance of the authorities. Many victims are threatened with further violence if they dare to reveal their oppression (e.g., row 5). Clearly, victims face challenges for breaking their silence, thereby not only hinting at the courage required to approach the first stage of police filing, but also the likely number of unreported cases that never enter the formal multi-stage process. The testimonies of the dowry murders (crimes that coincidentally have the highest percentage of suspect acquittal, Supplementary Table A21), illuminate the quantitative insights by demonstrating how real individuals are impacted.

While the testimonies in Table 7 highlight other themes beyond this article’s scope (e.g., the role of caste), the matched dataset facilitates an additional quantitative test for gender discrimination. In Table 8, the coefficients on female remain significant. Columns 1–4 in the top row show that the suspect implicated in a case registered by a woman has a significantly unlikely chance of being convicted or going to jail compared to those brought by men. Columns 5–8 restricts the sample to the court docket.

Table 7.

Balance Check II, Matched Cases and Corresponding Charges/Penal Codes [Identifying Information Redacted]

Table 8.

Impact of Complainant Gender on Conviction/Acquittal of Suspect in Case after Text Matching

Note: Controls include a numeric variable for a crime’s distance from a station and investigator rank. Columns 6 and 8 control for the rank of the judge. The top rows examine the effect of “female” post text matching, whereas the bottom rows exclude all cases with a VAW Penal Code so that the comparison is as far as possible restricted to generic or non-VAW cases. *p < 0.1; **p < 0.05; ***p < 0.01.

Matching cases of VAW between male and female complainants can be construed an odd comparison because, while topically similar, a man registering a case on behalf of a relative or friend might not have all the facts. Similarly, if a case of VAW is being lodged by a victim’s male family member or friend, it could imply that the victim has significant support or belongs to a particular class. Moreover, victims of VAW may be under trauma when providing testimonies, and so cases could be systematically different between men and women complainants who report such crime. As seen in row 3 of Table 7, the male and female complainant are both registering dowry cases; however, the man is emphasizing his economic losses whereas the woman is underscoring emotional and physical abuse.

For these reasons, one may want to make the comparison more parsimonious by matching on non-VAW (or generic) cases alone. In the bottom of Table 8, I re-run the algorithm to exclude VAWFootnote ⁵² such that the comparison is generally restricted to topics such as hit-and-runs, cheating, scooter theft, and burglary. The gender gap remains significant. Importantly, TIRM is likely an underestimate of discrimination. The approach understates the differential effort required by women to have reached the first stage or, say, norms about publicly coming forward. Moreover, women who register scooter theft (e.g., those that own such an asset or would even report it if stolen) may not be representative of women in Haryana society. The testimony upon which cases are matched might itself be gendered, for example, based on a lifetime of discrimination that generates differences in speech and word-choice. And so, while matching rests on several assumptions (Feder et al. Reference Feder, Keith, Manzoor, Pryzant, Sridhar, Wood-Doughty and Eisenstein2021), that discrepancies remain supports the preceding analyses in demonstrating that (gender) identity does have a bearing on criminal justice outcomes.