How Endogenous System Bias Can Distort Decision-Making in Criminal Justice Systems

Most judicial, regulatory, and administrative systems, at least formally, are concerned with the fairness and transparency of their decisions concerning the public. Fairness and transparency of criminal justice operations are critical to creating trust in the legal system and assuring people that the larger social system is legitimate and worthy of support. However, deviations from objective and fair decision-making can be concealed when key actors who are responsible for deciding outcomes in their organizations are also responsible for collecting, assembling, evaluating, and presenting the information on which their decisions are based. Under these conditions such systems are at risk of what we term “endogenous system bias,” where data are acquired and altered in ways to justify desired outcomes that are neither fair nor transparent. The purpose of this paper is to: (1) develop a general model of decision-making constraints that can produce endogenous system bias, (2) review research on endogenous system bias at two key decision stages in one institution: the criminal justice system and (3) conduct an empirical examination of the potential effect of endogenous system bias on law enforcement investigations and prosecutorial charging in a sample of criminal homicide cases in the USA. Implications for law, policy, and social scientific methodology are discussed.


Introduction
Judicial, regulatory, and administrative systems are at least formally concerned with the fairness and transparency of their decisions concerning the public. Fairness and transparency of criminal justice operations are critical to creating trust in the legal system and convincing people that the larger social system is legitimate and worthy of support (Blader & Tyler, 2003;Tyler & Huo, 2002). However, the decision-making integrity of such individuals and organizations is vulnerable to processes and forces that can distort their objectivity and impartiality. Equally important, deviations from objective and fair decision-making can be concealed when key actors who are responsible for deciding outcomes in their organizations are also responsible for collecting, assembling, evaluating, and presenting the information on which their decisions are based. Under these conditions, such systems are at risk of what we term "endogenous system bias," where data are acquired and altered in ways to justify desired outcomes that are neither fair nor transparent (because conflicting evidence has been dropped or falsified and supporting evidence has been distorted or fabricated). The purpose of this paper is to examine sources of endogenous system bias and assess the potential effect of this type of bias on evidence collection in the criminal justice system.
Notably, altering evidence to achieve or shape outcomes can occur in a broad range of private and public enterprises including criminal justice systems. For example, Chen et al. (2010), in a study of accounting oversight practices, found that company investment strategies differed when executive decision-makers and accounting controls intersected versus when they were independent of each other. 1 Of greater consequence, endogenous system bias also can affect how organizations meet their obligations to protect the public safety. Examples of decision-making failures facilitated by the biased collection of regulatory or oversite data occur with unfortunate regularity. These include such well-known examples of regulatory failure as the fatal crashes of two Boeing Airlines 737-Max planes in 2019 and 2020, and the pollution of Flint Michigan's drinking water with toxic levels of lead that began in 2014. In both cases, dangerous conditions were difficult to identify because organizational decision-makers (or actors reporting to them) were also responsible for monitoring the safety of their products or systems. In such instances, the resulting lack of transparency made it far more difficult to identify and remediate the underlying reasons for dangerous system failures.
The present analysis focuses on the operation of endogenous system bias in the criminal justice system in the USA, although many of the patterns we identify may also apply to decision-making in other arenas. In theory, the American criminal justice system is committed to principles of fairness, objectivity, and impartiality. In the USA, these principles are articulated in the constitution and have been interpreted and affirmed by centuries of court precedents (Legal Information Institute, n.d.; U.S. Courts, n.d.). Nevertheless, in the course of reconciling laws and principles with the specific circumstances of a given decision or case, criminal justice agents must make judgments and exercise considerable amounts of discretion, using "an authority conferred by law to act in certain conditions or situations in accordance with an official's or an official agency's own considered judgment and conscience" (Bureau of Justice Statistics, n.d.).
The exercise of discretion in the American criminal justice system may be sometimes difficult to assess because actors who are responsible for particular decisions (e.g., to arrest, to charge a person for a crime, to negotiate a plea or go to trial, to convict, to sentence, etc.) are also often the same actors responsible for acquiring, organizing and presenting at least some of the information on which their decisions are based. Under these circumstances, endogenous system bias can help to justify various criminal justice decisions though the social construction of evidentiary records to support desired and potentially biased outcomes. Further, evidence that is changed, falsified, or ignored at earlier stages in the criminal justice system likely conceals the presence of bias to decision-makers at later stages in the system. Similar challenges to the objectivity of criminal justice system decisions are also being recognized by European nations.For example, a report on Roma minorities in Europe by Fair Trails (December 2020) concluded that "Roma face structural discrimination throughout the criminal justice system from multiple actors-the police, judges, prosecutors, and even their own lawyers. (Fair Trails, p.1)" and such discriminatory views and attitudes can produce unfair outcomes for Roma defendants. A later study 2 on criminal justice proceeding in Belgium, Bulgaria, Greece, and Romania concluded "that justice does not apply neutrally and equally, and that criminal justice systems carry the biases of the wider society (Fair Trails, September 2022, p. 4)." More generally the EU Commission, in a recent report, on principles for national action plans against racism and racial discrimination pointed to the need address different forms of racism (conscious or unconscious, individual or structural) in all major EU institutions, including law enforcement and criminal justice systems.
Importantly, to the extent racism and other forms of prejudice influence of criminal justice decisions in different EU criminal justice systems, it is also possible that endogenous system bias may affect the collection of evidence in ways that conceal such influences. Finally, although the presence of endogenous system bias may vary across different organizations it is nevertheless most likely operated where the decision-making and evidence collection functions of a system intersect under the same authority.

3
Social Justice Research (2023) 36:192-224 That criminal justice systems in the USA and other nations sometimes deviate from the principles of fairness, and objectivity is not surprising in that they all operate within organizational, cognitive, social, and cultural constraints that can generate unfair, prejudiced, and obfuscated outcomes. Although the specific manifestations of such constraints vary across different nations and regions, some mix of these conditions are present in most countries and can affect their decision-makers. Critically, institutions whose organizational frameworks do not separate decision-making from the collection of data used to make those decisions will be less able to identify and remediate unfair and sometimes dangerous outcomes.
To describe and illustrate the concept of endogenous system bias, the paper is organized into three major sections. The first section develops the theoretical model for this concept broadly. We begin by developing a model for understanding endogenous system bias, describing three major types of system constraints that potentially can affect the collection of evidence upon which outcome decisions are based: (1) system constraints imposed on actors within organizations, (2) the cognitive and psychological constraints of decision-making actors within systems, and (3) the broader contextual activation of biases and prejudices that may distort actors' decisions.
The concept may best be understood by focusing on examples from a single system. To this end, in the second section we introduce the criminal justice system as a system in which endogenous bias may be particularly consequential. We examine relevant literature on how system, cognitive and contextual constraints may affect the decisions of criminal justice system actors and the evidence they collect (to support their decisions), focusing on two important stages of the criminal justice process: investigation and charging.
Finally, in the third section, we illustrate this concept with an empirical analysis of potential evidence of endogenous system bias in a particularly consequential justice system process: the death penalty. A large body of research in the USA has consistently found biases in death penalty outcomes. Endogenous system bias, however, points us toward an examination of the process-and specifically to whether there are biases in the purportedly objective evidence of such cases. This is a critical question because corrupted or altered evidence may not only bias outcome decisions, but at the same time give such decisions the appearance of fairness.
To examine the question of whether there is bias in the collection of evidence, the empirical analysis examines a particularly consequential justice system process; the death penalty. The analysis focuses on potential for bias at two key stages in the criminal justice process: (1) the collection of evidence by law enforcement officers on crimes they are investigating and (2) the formal charging of the criminal circumstances by prosecutors in cases they are prosecuting.
Since the empirical analysis does not have direct measures of the system, cognitive or contextual factors that may produce endogenous system bias the study draws on an extensive body of research that has found that the race and gender of homicide victims often affects criminal homicide sentencing. These studies typically controlled for a broad range of legally relevant and other relevant situational case factors while investigating the potential effect of demographic factors on sentence outcomes. Overall, these studies found that the race of homicide victims has 1 3 consistently shown an effect on the sentences received by homicide defendants across different regions in the USA. 3 The present study, using data from murder prosecutions in San Diego, County California covering the period 1978 to 1993, examines the impact of these two victim characteristics (i.e., race and gender), not on the sentence outcomes of the cases, but rather on (1) the evidence assembled by investigators about a crime, and (2) the formal criminal charges brought by prosecutors against the defendant in a case, i.e., whether the prosecutor charged special circumstances and sought the death penalty. At the time of the study, there were twenty-seven different special circumstances defined in California Penal Code § 190.2(a) (Appendix 1), at least one of which had to be formally charged by a prosecutor wishing to seek a death sentence.
In addition to the legal relevance of special circumstances, the regular collection and use of these circumstances provides for this analysis a set of well-defined indicators that examine potential patterns of bias in the collection and application special circumstance 'evidence.' The empirical analysis will test the hypotheses that: 1) specific demographic characteristics of homicide victims will affect the number of different special circumstances found by law enforcement investigators while controlling for legally relevant and other potentially relevant situational factors associated with the homicide case, and 2) specific demographic characteristics of homicide victims will affect the number of special circumstance charges brought by prosecutors while controlling for legally relevant and other potentially relevant situational factors associated with the homicide case.
Although the operation of arrest, prosecution and other criminal justice system decisions vary across nations, at least some national systems have decision-making and evidence collection/application responsibilities that intersect. The study's empirical analysis is intended to examine two instances where such overlapping responsibilities may bias the collection and application of evidence.

A General Model of Decision-Making Constraints that can Produce Endogenous System Bias
We begin by describing our concept of endogenous system bias through a review of three major types of decision-making constraints and their effect on the collection of evidence and organizational decision-making: system, cognitive and contextual level constraints.

System Constraints on Decision-Making
At the system level, criminal justice agencies, as with many public institutions, are subject to substantial resource constraints. Heavy workloads and associated time 1 3 Social Justice Research (2023) 36:192-224 pressures present additional challenges to making correct decisions and accurate judgments. Under such conditions, the effects of time pressure may cause decisionmakers to lower their standards for making accurate decisions, and as a consequence base their decisions on lower-quality or less information (Dror et al., 1999).
The organizational structure of the criminal justice system presents additional barriers to its capacity to function without bias. Like many government agencies, the criminal justice system is decentralized by geography, and also by function and level. This type of organizational structure can affect a system's capacity to share and analyze data, making it difficult for individual agencies and agents to share or obtain information, establish best data-driven practices, and as a consequence, enforce uniform procedural standards (see, e.g., Pierce & Griffith, 2005). What emerges is a situation in which there are often inadequate means to acquire the data criminal justice agencies need to perform systematic, independent evaluations.
At the individual level, significant contextual pressures can impact individuals' abilities to make undistorted judgments and decisions. Police officers must navigate employment environments that embed notoriously stringent institutional cultures (Skogan, 2008), which proscribe and prescribe behaviors that are not necessarily oriented toward the stated tenets of the justice system (Terrill et al., 2003(Terrill et al., , 2015. Moreover, police can operate in dangerous, tense, and changeable work environments that militate against the exercise of neutral judgment (Regehr & LeBlanc, 2017;Shane, 2010).

Cognitive and Psychological Constraints on Decision-Making
The constraints of the criminal justice system and contextual pressures on decisionmakers are compounded by well-documented cognitive and psychological limitations on human decision-making. Research in evolutionary psychology and social cognition reveals a myriad of ways in which human perception and mental processing shape the accuracy of our impressions, interpretations, and judgments (Ehrlinger et al., 2016;Tversky & Kahneman, 1974). Haselton et al., (2016: 968) define cognitive bias cases as "cases in which human cognition reliably produces representations that are systematically distorted compared to some aspect of objective reality." From this perspective, truth and objectivity are not necessarily primary pursuits of our cognition. Instead, powerful, ancient incentives such as reducing risk and acquiring social esteem-that have benefits for survival or reproduction-also shape human cognition and the cognitive devices by which people interpret perceptions and make judgments (Haselton et al, 2009(Haselton et al, , 2016Pinker, 2005). Even values, which might represent good-faith judgments about empirical reality, can arise within an incentive environment where their objective accuracy is often in tension with their social desirability (Haselton et al., 2009).
The limitations and tendencies of human cognition can produce cognitive bias or error in decision-making. These types of errors have been studied and incorporated into other disciplines, including social and evolutionary psychology, behavioral economics, and game theory (Glaeser, 2004;Kahneman, 2003).

3
Some cognitive errors of particular importance to criminal justice decision-making operate something like logical fallacies in the interpretation of information or stimuli. For example, attribution error refers to the tendency to attribute the behavior of others to their intrinsic nature or personality while discounting the causal influence of external or situational factors (Gawronski, 2007;Ross, 2018). Anchoring is a type of cognitive error wherein a person unduly relies on a single data point when making decisions; often relying on the first impression or piece of information acquired (Englich et al., 2006;Gilovich, 2007;Tversky & Kahneman, 1974).
Finally, some cognitive errors can be understood as modes of motivated reasoning, wherein people disproportionately select, construct, and interpret information and arguments that support or are consistent with their preexisting and preferred conclusions (Kunda, 1990;Pandelaere, 2007). For example, confirmation bias refers to the tendency to over-value information which accords with one's prior belief or desired conclusion and discount contradictory evidence (Casad, 2007;Devine et al., 1990;Westen et al., 2006); belief perseverance refers to the tendency to maintain a belief even when presented with compelling evidence that the belief is false (Anderson, 2007;Tetlock, 1983); and system justification refers to motivated defense and affirmation of the status quo with respect to the social, political, or economic system in which one participates (Jost & Liviatan, 2007;Jost et al., 2018).
Moreover, research on human decision-making biases from the psychological, social psychological and behavioral economics sciences has begun to receive support from the cognitive and neurosciences (see generally Sapolsky, 2019). As an example, studies have found increased activity in the dorsal anterior cingulate cortex (dACC) and the dorsolateral prefrontal cortex (DLPFC) when subjects are confronted with other-race faces. Variation in activation of the dACC has been found to correlate with White subjects' pro-White implicit association test (IAT) scores (Kubota et al., 2012), suggesting a greater cognitive effort is required on the part of those with implicit racial preferences to prevent those preferences from shaping their behavior.

Contextual Activation of Bias and Prejudice on Decision-Making
Although the neurological and cognitive structures which facilitate these biases and errors are present from birth, they are malleable and subject to contextual activation (Sapolsky, 2019). Hence, the social context provides the categories, distinctions, and associations through which these biological tendencies express themselves.
In this respect, human decision-making can be seen as a product of not only our cognitive makeup but also our prior experience and memories and the social and cultural context within which decisions are made. Of particular relevance for criminal justice decision-makers are the social contexts and individual beliefs associated with race, gender, ethnicity or other factors that may elicit fear or prejudice. Race, for example, which has been found to be a significant predictor of criminal justice outcomes virtually across the board, has a deep and longstanding association between fear of crime and socio-political mobilization that is animated by linking racial stereotypes and criminality (for in-depth analyses of these dynamics, see Hagan, 2010;Simon, 2007).
Public fears and associations can be intensified under the operating conditions of a criminal justice system characterized by high workplace expectations, high workloads, resource scarcity, and high stakes outcomes. As a result, criminal justice work environments can often amplify effects of cognitive biases and distort the objectivity of decision-makers in the system and what they collect as evidence.
In addition to the pressures of the work environment, actors in police and court settings respond to professional work incentives. Chief among these are the metrics by which their performance is measured: police according to clearance rates (arrests) and prosecutors according to successful prosecutions via guilty verdicts or pleas. External incentives, such as political/public pressure resulting from fear of crime (both justified and not), can similarly shape the way criminal justice agents go about their work. The combination of the pressures and incentives present in criminal justice work environment, along with recognized potential individual (conscious and unconscious) biases, means that criminal justice decision-makers (or actors in other organizations with at least some of these conditions) may be especially susceptible to either developing erroneous evidence records, being the consumers of such evidence from earlier stages in the process or being parties to both processes.

Research on Endogenous System Bias at Two Key Decision Points in One Institution: The Criminal Justice System
Although our conception of endogenous system bias is relevant for a wide variety of types of systems, to better understand the implications of the concept, we focus on one system as a consequential exemplar: the system tasked with administering legal justice in the USA. Within this system, we focus on critical stages in the decisionmaking process: criminal investigations and prosecutorial charging decisions.

Endogenous System Bias in Criminal Investigations
The investigation phase of the criminal justice process develops evidence upon which each subsequent stage builds. Hence, it represents an important opportunity to identify (or introduce) errors and prevent biases from corrupting the evidentiary record. While quantitative data on many aspects of police investigations are scarce, significant data are available on police-citizen contacts, including stops and searches that are strongly suggestive of systemic biases in judgment and potentially how evidence is collected.
Research consistently finds that African-Americans in general and young Black men in particular experience more police contact than do Whites. African-Americans are significantly more likely to be stopped (Bureau of Justice Statistics, 2013; Crutchfield et al., 2012), searched (Baumgartner et al., 2017;Bureau of Justice Statistics, 2013;Rojek et al., 2012), and arrested (Baumgartner et al., 2017;Kochel et al., 2011;Lytle, 2014;Mitchell & Caudy, 2015) by police than similarly situated Whites. The difference in rates of searches is especially concerning in light of 1 3 evidence that frisks of White suspects are more likely to uncover contraband (Jones-Brown et al., 2010).
Many of these disparities cannot be accounted for by variations in offending rates between Whites and minorities. Mitchell and Caudy (2015) found that "at ages 17, 22, and 27 African-Americans' odds of drug arrest are approximately 13, 83, and 235% greater than Whites, respectively" (p. 306-307), despite a finding that "African-Americans and Hispanics reported statistically lower rates of drug offending [compared to Whites] on nearly every measure of drug offending." Although these and similar studies focused on criminal justice outcomes, they also created a disparate evidentiary history for Whites, African-Americans and Hispanics that likely affected later encounters with the criminal justice system.
Recent research on implicit bias indicates the important (if unconscious) role of individual agents in fostering racial disparities in criminal justice outcomes (e.g., Hewstone, 2002;Quillian, 2008). A variety of studies suggest implicit biases (e.g., Banaji & Greenwald, 2013;Greenwald et al., 1998) and implicit animus (e.g., Payne et al., 2005) are relatively common and associated with a wide variety of attitudes and behaviors. This includes attitudes and behaviors relevant to criminal justice system processes. Among the general public, implicit racial animus appears associated with perceptions of increasing crime, fear and anger at the prospect of victimization, genetic attributions for racial disparities in criminal behavior, and support for punitive crime policies (Drakulich & Siller, 2015;Drakulich, 2015aDrakulich, , 2015b, and research has found that individuals have an easier time linking weapons and street crime to Black faces (Banaji & Greenwald, 2013). These same forces may also directly affect actors in the criminal justice system. At least one study of police officers found that priming officers to think about crime directed them toward Black faces, and that they tended to rate Black faces-especially stereotypically Black faces-as more likely to be criminal (Eberhardt et al., 2004). This suggests that implicit biases of both the public and criminal justice actors may affect the type of evidence provided by citizens and collected by law enforcement.
However, more directly with regard to collection of evidence, there is an emerging exchange among law enforcement professionals regarding the role of cognitive errors in police investigations and the construction of evidence (e.g., Heuer, 2007;Pinizzotto et al., 2004;Rossmo, 2006aRossmo, , 2006bRossmo, , 2009. Of primary concern for these practitioners are the effects of anchoring, confirmation bias, and belief perseverance. Together these habits create "tunnel vision" among investigators wherein their initial impressions of a case exert disproportionate influence on their developing beliefs (anchoring) of a case by seeking confirmatory evidence more vigorously, weighing initial evidence more heavily than potentially exonerative evidence, and allowing the dominant theory of a case to persist even in the presence of contradictory information (belief perseverance) (Findley & Scott, 2006;Martin, 2002;Petersen, 2017;Rassin, 2018;Thompson & Schumann, 1987). 4

3
Social Justice Research (2023) 36:192-224 Research also suggests criminal investigations can be significantly influenced even by subtle changes in the investigative process and context. For example, Lidén et al. (2018) found that knowledge that a suspect had been apprehended caused subjects (a law enforcement sample and a student sample) to interrogate the suspect in a more guilt presumptive manner and rate him or her as less trustworthy. Canter et al. (2012) found that knowledge of the identity and position of a target suspect within a line-up by a line-up administrator (compared with no such knowledge) doubled the likelihood that a witness would identify the target suspect.
Similarly, Bridges and Steen found that differential attributions about the causes of crime by probation officers produced "pronounced differences in officers' attributions about the causes of crime by white versus minority youths" (1998: 554). These differences contributed to racial disparities in probation officers' assessments of defendants' risk of reoffending and consequently the officers' sentencing recommendations.
Research also documents the role of cognitive biases in the development of forensic evidence. Edmond et al. (2015) documented the impact of what the authors refer to as "contextual bias." This type of bias can arise when forensic practitioners are (1) exposed to erroneous, potentially biasing information about the case which in turn may make them (2) liable to spread their "cognitively contaminated" interpretations to other witnesses or professionals in the criminal justice system. Similarly, Dror et al. (2006) studied the effects of contextual information on fingerprint experts. Their analysis found that the suggestion that two sets of prints were not a match caused a majority of experts to contradict their own prior determinations that the prints were a match.
More generally, in a review of research on the impact cognitive biases have on forensic science practitioners, Cooper and Meterko (2019) found substantial evidence of confirmation bias in forensic analysis, especially in studies wherein subjects were presented with information concerning the suspect or crime scenario. Moreover, in a study of 403 experienced forensic examiners regarding their views of cognitive bias and its role in their profession, Kukucka et al., (2017: 456) found that many of the practitioner interviewees "have only a limited appreciation of cognitive bias or see themselves as impervious to it," despite acknowledging the role of cognitive bias in other disciplines.

Endogenous System Bias in Prosecutorial Charging and Decision-Making
In the American criminal justice system, prosecutors are expected to fill a dual role in the criminal justice system, acting simultaneously as advocates of a state or federal jurisdiction and as "ministers of justice." According to the American Bar Association, "This responsibility carries with it specific obligations to see that the defendant is accorded procedural justice, that guilt is decided upon the basis of sufficient evidence and that special precautions are taken to prevent and to rectify the conviction of innocent persons" (American Bar Association, 2021). However, the ethical demands of ensuring decisions are based on sufficient evidence can come into tension with the incentives emanating from the imperative of advocacy.
For example, as Medwed (2012) notes, prosecutors must maintain good working relationships with the law enforcement officers who initially collect evidence and refer suspects to them for charging. Prosecutors may encounter questionable evidence, contradictory witness accounts, or confessions that are possibly coerced. Proper and ethical courses of action would necessarily involve second-guessing the integrity of the evidence. However, in so doing they would be calling into question the competence and integrity of the police officers as well, upon whom the prosecutor must be able to rely in future cases. This is no small matter. Prosecutors who develop a reputation for being difficult or hard on police are often punished through the withholding of necessary information about police investigations (Medwed, 2012: 23-24). Medwed (2012:38) further argues that the tension arising between prosecutors' roles as ministers of justice and advocates renders them susceptible to cognitive dissonance avoidance techniques. Cognitive dissonance is uncomfortable, and people will often adjust their beliefs to be more congruent with their actions in order to alleviate the discomfort (Acharya et al., 2018;Colosio et al., 2017;Festinger, 1957;Izuma et al., 2010). These mental adjustments may take the form of common cognitive errors, such as confirmation bias and motivated reasoning that produce adjustments between beliefs and actions when individuals seek to avoid cognitive dissonance (Acharya et al., 2018: 409).
Thus, in order to retain functional relationships with the police who ostensibly believe they have arrested a likely suspect, prosecutors may adjust their own evaluations and accept potentially biased forms of evidence. Similarly, political pressures to move a high-profile case forward and obtain convictions may exert pressure on the quality and presentation of evidence acquired for cases. Equally important, prosecutors are measured and evaluated by obtaining outcomes on behalf of the state.
Evidence may not only be acquired or constructed to advance desired outcomes, but it is also possible to hide exculpatory information to attain such an outcome. As an example, Medwed (2012) notes that the US Supreme Court's 1963 holding in Brady v. Maryland (373 U.S. 83) established that a prosecutor's failure to turn over to the defendant material evidence in the prosecutor's possession was a violation of due process. However, the 1963 decision also granted prosecutors initial discretion to determine whether potentially exculpatory evidence is material to the defendant's case. Evidence deemed immaterial by the district attorney's office is not turned over to the defense and accordingly is often unknown to defense counsel and the court. The discovery of potentially exculpatory evidence following a decision to charge a suspect creates a tension between advocacy and justice; discounting the importance or reliability of that evidence can help alleviate such tension.
Work by Spohn et al. (2001), Steffensmeier et al. (1998), and Steffensmeier and Painter-Davis (2017) developed and applied a conceptual framework that examined how prosecutorial charging outcomes and sentencing decisions engage "focal concerns" which can function as a form of perceptual shorthand in assessing such issues as convictability, dangerousness, and blameworthiness. This type of perceptual shorthand can help criminal justice actors manage heavy workloads but also can help actors reduce cognitive dissonance about the evidence they obtain and the decisions they make. Ulmer (2019) found that local specific "focal concerns" can evolve to align with local contexts and the predispositions of actors within those contexts. This area of research also found that charging and sentencing outcomes are related to the intersectionality of various racial and demographic combinations of the victims and defendants.
Prosecutorial decisions concerning the charging and prosecution of defendants also appear vulnerable to bias. Wu (2016) conducted a meta-analysis of 26 studies and found significant race effects on the chances of being charged and fully prosecuted, with the most pronounced minority disadvantages occurring at the charging stage. Smith and Levinson (2012) examined the impact of implicit racial bias on the exercise of prosecutorial discretion. Reviewing previous empirical studies, they found that prosecutors were less likely to charge White suspects than Black suspects in similar cases.
Another line of research examines disparities from a cumulative perspective, and the effect of decisions made at earlier stages in the process on later outcomes. This approach views the criminal justice system as a process wherein any errors or biases occurring in any given stage can be carried forward and generally be unquestioned in subsequent stages. In a study of racial variation in the prosecution process Stolzenberg et al. (2013) found that Black defendants were 42% more likely to receive a severe sanction, even when controlling for prior record and other legal and extralegal factors. In an exploration of cumulative disadvantage in prosecution and sentencing in New York County (i.e., Manhattan), Kutateladze et al. (2014) found variation across discretion/decision points as well as offense categories that worked against Black and Hispanic defendants. Focusing on mandatory sentencing, Starr and Rehavi (2013) found that racial disparities were introduced during several stages of the criminal justice process, and that much of the disparities stemmed from prosecutors' charging choices, specifically their decisions to charge defendants with offenses that carried mandatory minimum sentences.

An Empirical Examination of the Potential Effect of Endogenous System Bias on Investigations and Prosecutorial Charging in Criminal Homicide Cases
The literature review provides consistent evidence that cognitive biases and social context can affect the collection, presentation, and evaluation of evidence in the criminal justice system and that such biases can create a corrupted evidentiary record that is difficult to spot by later actors in the system. However, these are difficult questions to assess because of the paucity of comprehensive data on criminal justice decision-making and outcomes, and as discussed above, because of potentially biased evidentiary records. Difficult though they may be, it is not always impossible to assess how the books may be cooked.
Homicide cases can provide one potential exception to this rule because more information is typically collected in such cases. In addition, researchers concerned with possible bias in capital punishment cases often spend considerable time collecting paper or electronic archived records on homicide cases. Therefore, this type of research sometimes acquires enough data to identify patterns of biased evidence construction (see, e.g., Pierce et al., 2014;Radelet & Pierce, 1985). Radelet and Pierce (1985) linked Supplementary Homicide Report data from police agencies on homicides to subsequent court data on the same cases to assess whether changes in the felony circumstance designation of a case were related to race. Pierce et al. (2014) were able to examine whether the volume of evidence provided by prosecutors to defense attorneys was related to race because the county court's information system recorded the number of pages of evidence provided by prosecutors to defense attorneys through the discovery process. Figure 1 maps, in a simplified manner, the major stages of the criminal justice process wherein the exercises of judgment, and the development of evidence upon which judgements are made are at risk of being biased. The empirical analysis focuses on evidence collected in the first two stages of the process; the investigation and charging stages. Figure 1 also shows the potential downstream impact of bias at earlier stages on decisions made at later stages (e.g., conviction, sentencing and appellate review as shown in Fig. 1) because biased evidentiary records can be carried forward and taken as unbiased 'facts' upon which subsequent decisions are grounded. Moreover, the accumulation of imperfect evidence also bestows legitimacy on subsequent decisions and makes it more difficult to identify significant errors in the criminal justice process.
To zero in on this question, we use data from one such study that were collected for a capital post-conviction case in San Diego County covering the period 1978 to 1993 (Shatz et al., 2020). Extensive data on various aspects of criminal homicide prosecutions were collected to determine if there was evidence of racial bias in the capital punishment decisions. A major focus of that study was the charging decisions by the San Diego County District Attorney's Office. For the present analysis, data from the study allowed us to examine two different stages in the collection and application of evidence: the compilation and recording of special circumstance evidence by police and prosecutors, and the subsequent charging of special circumstances by prosecutors. These two stages are important because charging a special circumstance is a critical decision in the California criminal justice process and because the death penalty can only be sought if a special circumstance is formally charged by the prosecutor.
The empirical analysis will specifically examine whether the race and gender of homicide victims in San Diego County murder prosecutions are determinants: (1) of the number of different instances of special circumstances evidence collected by law enforcement investigators controlling on legally and situationally relevant factors of each case and (2) of the number of special circumstances charged by prosecutors while controlling for legally and situationally relevant factors of each case.

Data Sources
The primary sources of the data form the San Diego study were: (a) copies of charging documents provided by the District Attorney for cases in which a violation of Penal Code Sect. 187(a) (murder) was charged during the study period, and (b) copies of the pre-sentence reports ("PSR"s) for defendants in those cases resulting in a conviction. Charging document and PSR data were supplemented with information from the State of California Department of Justice Willful Homicide Charts, the Federal Bureau of Investigation Supplementary Homicide Reports, appellate court opinions, and newspaper accounts of the crimes and prosecutions.
Data were collected on 1,081 cases for the years 1978-1993 in which an adult defendant was charged with a violation of Penal Code Sect. 187(a) and was convicted of a homicide (see Shatz et al., 2020). For the purposes of this analysis the sample was further restricted to cases where there was a White, Latinx, or Black defendant and where the homicide case involved at least one victim who was White, Latinx, or Black. This resulted in a final sample of 996 cases.
Each case was coded for factors concerning the defendant, the victim, the crime, and the prosecution. Since our focus was on the death penalty, we coded each case for the presence or absence of the special circumstances as set forth in Penal Code § 190.2(a) that would have made the defendant eligible for a death sentence. A special circumstance was coded as present "if the circumstance was found by a fact-finder or admitted by the defendant or if the facts of the case were such that a reasonable fact-finder could have found the circumstance true beyond a reasonable doubt" (Shatz et al., 2020(Shatz et al., , p.1085). The authors treated as controlling a fact-finder's determination that no special circumstance was proved. Under these standards, a total of 447 of the 996 cases (44.9 percent) contained sufficient evidence of the presence of one or more special circumstances (hereinafter, "special circumstances cases").
Data were also coded for other aspects of the crime: whether there was a vulnerable victim (a child or elder); whether the victim relationship involved domestic violence or whether the victim knew the victim but the crime was not domestic violence; whether the role played by the defendant in the killing was as a principal or non-killing accomplice or co-conspirator; whether a firearm was used; and whether  the defendant caused injuries to persons other than the homicide victim. Also measured were two statutory aggravating factors: prior felony convictions and other proved or provable crimes involving violence or the threat of violence. Finally, data on non-statutory factors about the defendant that may have influenced the charging decision were obtained: whether the defendant was on parole or probation at the time of the crime, the relationship of the victim to the suspected offender, and whether the defendant was a gang member at the time of the crime. Lastly, data was coded for the race or ethnicity and gender of the defendant and the victim(s).

The Impact of Victim Race and Gender on the Collection of Investigative Evidence
California Penal Code § 190.2(a) specifies the special circumstances which, if admitted by the defendant or found true at trial, subject a defendant to a sentence of life without parole or death. In the study sample of 996 cases, there were 447 special circumstances cases (44.9%) in which one or more different special circumstances appeared. During the period of our study, there were 27 separately enumerated special circumstances until June 5, 1990, and 29 separately enumerated special circumstances thereafter. 5 These data allow us to examine whether findings of special circumstances are associated with either race and/or gender of victims. Table 1 presents the number of special circumstances by the race and gender of the victims. Race is coded into categories of those cases with one or more White victims versus homicide cases with no White victims, but at least one Black or Hispanic victim. These categories are labeled as White and non-White victim homicide cases. Homicide cases are classified as female victim homicides if the case involved at least one female homicide victim. The final classification results in four categories: 1 = White Female, 2 = Non-White Female, 3 = White Male, and 4 = Non-White Male homicides.
As Table 1 shows, no sufficient special circumstance evidence was present in 55.1% of the 996 cases, one special circumstance was present in 25.6% of the cases, two special circumstances were present in 12.0% of the cases, and 3 or more different special circumstances were present in only 7.2% of the cases. Table 1 also reveals significant differences in the number of special circumstances present across the four different race and gender victim groups. Specifically, in 33.6% of the White Female victim homicide cases, two or more special circumstances were present, versus 23.5% for White Male victim cases, 12.0% for Non-White Female cases and 11.6% for Non-White Male victim cases.
Are the race and gender victim disparities in special circumstances cases partially a function of differential law enforcement investigatory efforts, or are they mainly a function of variations in the types of homicides associated with different categories of the victims' race and gender? To address this question, we conducted a Negative Binomial regression analysis of factors affecting the number of different types of special circumstances present in the 996 sample cases (Table 2). 6 The analysis included three race and gender of victim dichotomous variables, one race of suspect dichotomous variable, and 16 dichotomous indicator controls for the type of homicide case. The control variables included whether or not the defendant had prior adult or juvenile felonies, committed prior crimes of violence, was on probation or parole at the time of the homicide, was identified as in a gang at any time, was a principal in the homicide case, or was female. Additional control variables included: whether the case involved child or elderly victims, whether a firearm or drugs/alcohol were involved, the type of relationship between the suspect and victim(s), the presence of non-homicide victims/injuries associated with the crime, and whether the crime involved multiple homicide victims. Appendix 2 presents descriptive statistics for the dependent and control variables used in the multivariate analysis of the number of special circumstances present in the 996 criminal homicide cases.
For this analysis, we use a negative binomial model for counts of the number of special circumstances present in criminal homicide cases. Notably, as shown in Table 2, the Negative Binomial analysis found that the race and gender of victims continued to be significant predictors of the of the number of special circumstances present, even after controlling on the 16 other characteristics of the homicide cases. However, the race of the suspect was not a significant predictor. The results indicate that facts as developed by the police and/or reported by the probation officer appear biased toward finding homicides more aggravated in cases with White and/or female victims. This is key because biased evidence at earlier stages in the criminal justice system (in this case, the production of special circumstance evidence) is likely to carry over into later stages in the process. Importantly, actors at later stages in the process will have little way of knowing if the 'evidence' provided to them from an earlier stage is biased.
The Negative Binomial analysis confirms that some of the characteristics of homicide cases were statistically significant predictors of the number of specials present: prior crimes of violence by the defendant, the role of the defendant in the crime, whether there was child, elder or multiple victims of the crime, and the relationship between the victim and the defendant and the presence of a firearm. What is particularly notable, however, is even with the strong control variables, the race and gender of victims are also statistically significant predictors of the number of special circumstances present in PSRs or court records. Specifically, cases with White female or White male victims have significantly more special circumstances present, even with the control variables in the equation.

The Impact of Race and Gender on Prosecutorial Charging
The second question we examine is, among those 447 special circumstances cases, whether decisions of prosecutors to charge special circumstances are correlated with race and gender of the victim. To examine this question, we again classified the race and gender of victims into four categories (as described above). Table 3 presents the number of special circumstances charged by prosecutors by the race and gender of the victims. A special circumstance, if charged, not only makes a defendant eligible for a death sentence, but it is also an "aggravating factor" that can be considered by the jury at the penalty phase of the trial. Table 3 shows two clear patterns. First, examination of the total number of charges brought by prosecutors indicates that in only 23.1% (i.e., 100.0%-72.9%) of the special circumstance cases were defendants actually charged with a special circumstance by prosecutors. Second, Table 3 shows significant variations in the number of special circumstances charged by prosecutors by the race and gender of victims. Overall, 33.7% of the White female victim homicide cases were charged with two or more special circumstances versus only 23.1% for Non-White female victim cases, 12.3% White male victim cases, and 5.2% for Non-White male victim cases. The divergent patterns of prosecutorial charging by the race and gender of victims are equally clear (but reversed) for other categories of victim demographics where no special circumstances were charged.
These results raise the next question: whether the race and gender of victim disparities in prosecutorial charging of special circumstances are (at least partially) a function of differential law enforcement investigatory efforts and/or prosecutorial biases or whether they are mainly a function of variations in the type of homicide associated with different demographic categories of the homicide victims. To address this question, we use a Poisson model for counts of number of special circumstances charged by prosecutors among 447 cases where at least one special circumstance was present (see Table 4). 7 As in the previous analysis of the number of special circumstances present in the 996 criminal homicides cases, this analysis includes three race and gender of victim dichotomous variables, one race of defendant dichotomous variable, and the dichotomous indicator controls on type of homicide case. Also, like in the previous analysis, we included 15 dichotomous indicator controls for types of homicide case. We also included eight categories of special circumstances cases as controls on the type of homicide case. We aggregated the special circumstances into eight groups: Table 3 Number of specials circumstances charged by race and gender of victims in criminal homicide cases with at least one special circumstance present Pearson Chi-Square59.856 df = 9 sig. = 0.000 3 cells (18.8%) have expected count less than 5. The minimum expected count is 2.01 Victim race and gender Number of special circumstances charged by prosecutors (1) multiple murders; (2) sexual assault; (3) torture and kidnapping; (4) theft felonies; (5) designated victims; (6) financial gain; (7) lying in wait; and (8) miscellaneous. 8 These represent especially important controls for the types of homicide cases because they represent the evidence, which we would expect to be predictive of special circumstances formally charged by prosecutors. Appendix 3 presents descriptive statistics for each of the special circumstances charged by prosecutors in cases with at least one special circumstance present. The analysis was conducted using a Poisson model, and since the dispersion parameter was 1.09 and not significant at p = 0.093. 9 The Poisson analysis confirms, as would be expected, that the special circumstance present are indeed important predictors of the number of special circumstances actually charged by prosecutors, with six out of eight of these indicators showing a significant effect on specials charged. Among the other characteristics of homicide cases, the relationship between the victim and the defendant and the presence of a firearm are also important predictors of specials charged by prosecutors. What is particularly notable, however, is even with the strong evidentiary control variables in this analysis-the presence of most of the special circumstances,-the race and gender of victims continue to be significant predictors of prosecutorial charging decisions. Thus, the analysis of special circumstance evidence the analysis of prosecutorial charging decisions indicates that at two key stages in the criminal justice decision-making process-evidence collection and later prosecutorial charging decisions-the race and gender of victims remain important predictors of criminal justice outcomes controlling for a broad range of case characterizes.
One final note. Table 4 displays the distribution of cases across the five main stages of the criminal justice system in death penalty cases, broken down by the race and gender of the victim, and Fig. 2 displays these results in the form of a graph. The distribution of cases by race and gender of victim changes dramatically as the case proceeds. For example, we can see that homicide cases with at least one White female victim represent only 15 percent of the first-degree murder cases but grow to 36 percent of the cases in which a special circumstance was charged, 48 percent of the cases where a notice to seek the death penalty was filed, and 60 percent of the cases that ended with a death sentence. Conversely, cases with non-White male victims constituted 40.5 percent of the first-degree murder cases, and only 15 percent of the cases that ended with a death sentence. Some cases are prosecuted more vigorously and some less vigorously, and these differences correspond with the race and gender of the homicide victims. 10 It should also be noted that these data start at a point in the criminal justice process where homicide cases are already being adjudicated and both defendants and victims are known. In such cases, as we have seen, the characteristics of a victim appear potentially important factors in how cases are developed and charged against defendants. The importance of victim characteristics is typically not relevant in a very board category of criminal justice activity where there is no victim or the victim is unknown (e.g., drug offenses, police stops). Here research has found that the race and ethnicity of a suspect are important factors in law enforcement decisions (e.g., Pierson et al., 2020), and that there are compounding effects of racial biases in contact with the various stages of the criminal justice system (e.g., Baumgartner et al., 2017;Crutchfield et al., 2012;Drakulich & Rodriguez-Whitney, 2018).
In the end, among the 447 cases from San Diego in which the District Attorney could have charged one or more special circumstances, a death sentence was imposed in 13.2% of the cases with a White victim and a Black or Latinx defendant (10 ÷ 76), but only 2.7% of the time in all other cases (10 ÷ 371). None of the 36 cases with Black or Latinx victims and White defendants resulted in a death 10 A similar pattern of a change in the distribution of case by race has been documented in the research conducted in Louisiana (Lyman, Baumgartner, & Pierce, 2021), sentence. We conclude that these disparities have their roots in endogenous system bias, which begins to corrupt decision-making long before the evidence of a special circumstance is constructed.

Conclusions
The criminal justice system in the USA is only one example of the numerous systems in which endogenous system bias leads to distorted decision-making. Endogenous bias can arise in any system wherein actors who are responsible for making key decisions are also at least partially responsible for producing information and data that forms the basis for their decisions. We conclude that evidence produced in criminal justice cases may be in part a function of human cognitive and organizational limitations and biases, and not necessarily an independent representation of "objective" facts. This can be a particularly serious problem to identify in systems like criminal justice that have multiple successive decision points because as each decision point is reached and the case proceeds to subsequent decisions, it becomes increasingly difficult to shine light on prior bias if the evidentiary record itself is biased. For example, proportionality review that searches for comparatively excessive sentences is doomed to fail if only cases in which a death sentence has been imposed are included in the examination. These same factors represent important considerations for researchers examining questions of potential bias based on data derived from later stages in the criminal justice decision process (e.g., cases identified as eligible for a death sentence).
Only when we examine the entire series of decisions that culminate with a specific outcome (e.g., a death sentence) can we see how criminal justice work environments can (often unwittingly) amplify effects of cognitive biases and distort the objectivity of decision-makers in the system. The bias gets more difficult to spot as one moves from the origins of the decision (the crime) to the end (sentence). Because such a process may result in corrupted and biased forms of evidence, it becomes increasingly difficult to see biased decisions that are made at earlier stages closer to the time of the crime. Despite the difficulties, we were able to study the operation of the criminal justice system at two stages of a potentially capital case: investigation and charging. We found what endogenous system theory would have predicted: bias at the investigative stage amplified by bias at the charging stage.
In formal research studies, endogenous system bias is a special case of confounding or omitted variable bias. Confounding biases originate from common causes and "results from failing to control for a common cause variable that one should have controlled for" (Elwert & Winship, 2014, p. 32). Omitted variables are endogenous to a system when they arise from actions by actors within an organization regarding evidence they assemble, and when their decisions based on such evidence. This type of bias is particularly challenging to identify and control for in organizations because it may not only be unobserved (in methodological terms), but because they actually become unobservable (at least without significant effort to identify their presence).
Addressing the inaccuracies introduced by endogenous system bias is not a simple task. Clearly better and more objective data are needed to ensure that public policy decisions in general, and in the criminal justice system are more valid, accurate, and fair. Knowledge of the presence of endogenous system bias is the first step in reducing it, because the sources of this type of bias(es) are often (but not always) introduced without conscious intent or awareness. However, understanding the impact of endogenous system bias is the first step to mitigating it, and a key to ensuring transparency and fairness in criminal justice systems and other institutions responsible for the safety and well-being of the public.

Appendix 1
Special circumstances contained in Penal Code § 190.2(a) There were two slightly different versions of § 190.2(a) during our study period. From Nov. 7, 1978 (the enactment of the California Death Penalty Law) through June 5, 1990, § 190.2(a) listed the following special circumstances: (1) The murder was intentional and carried out for financial gain.
(2) The defendant was previously convicted of murder in the first or second degree.
For the purpose of this paragraph, an offense committed in another jurisdiction which if committed in California would be punishable as first or second degree murder shall be deemed murder in the first or second degree. (3) The defendant has in this proceeding been convicted of more than one offense of murder in the first or second degree. (4) The murder was committed by means of a destructive device, bomb, or explosive planted, hidden or concealed in any place, area, dwelling, building or structure, and the defendant knew or reasonably should have known that his act or acts would create a great risk of death to a human being or human beings. (5) The murder was committed for the purpose of avoiding or preventing a lawful arrest or to perfect or attempt to perfect an escape from lawful custody. (6) The murder was committed by means of a destructive device, bomb, or explosive that the defendant mailed or delivered, attempted to mail or deliver, or caused to be mailed or delivered and the defendant knew or reasonably should have known that his act or acts would create a great risk of death to one or more human beings. (7) The victim was a peace officer as defined in Section 830. 1, 830.2, 830.3, 830.31, 830.35, 830.36, 830.4, 830.5, 830.5a, 830.6, 830.10, 830.11 or 830.12, who, while engaged in the course of the performance of his duties, was intentionally killed, and such defendant knew or reasonably should have known that such victim was a peace officer engaged in the performance of his duties; or the victim was a peace officer as defined in the above enumerated sections of the Penal Code, or a former peace officer under any of such sections, and was intentionally killed in retaliation for the performance of his official duties. (8) The victim was a federal law enforcement officer or agent, who, while engaged in the course of the performance of his duties was intentionally killed, and such defendant knew or reasonably should have known that such victim was a federal law enforcement officer or agent, engaged in the performance of his duties; or the victim was a federal law enforcement officer or agent, and was intentionally killed in retaliation for the performance of his official duties. (9) The victim was a fireman as defined in Section 245.1, who while engaged in the course of the performance of his duties was intentionally killed, and such defendant knew or reasonably should have known that such victim was a fireman engaged in the performance of his duties. (10) The victim was a witness to a crime who was intentionally killed for the purpose of preventing his testimony in any criminal proceeding, and the killing was not committed during the commission, or attempted commission of the crime to which he was a witness; or the victim was a witness to a crime and was intentionally killed in retaliation for his testimony in any criminal proceeding. (11) The victim was a prosecutor or assistant prosecutor or a former prosecutor or assistant prosecutor of any local or state prosecutor's office in this state or any other state, or of a federal prosecutor's office and the murder was carried out in retaliation for or to prevent the performance of the victim's official duties. (12) The victim was a judge or former judge of any court of record in the local, state or federal system in the State of California or in any other state of the United States and the murder was carried out in retaliation for or to prevent the performance of the victim's official duties. (13) The victim was an elected or appointed official or former official of the Federal Government, a local or State government of California, or of any local or state government of any other state in the USA and the killing was intentionally carried out in retaliation for or to prevent the performance of the victim's official duties. (14) The murder was especially heinous, atrocious, or cruel, manifesting exceptional depravity. As utilized in this section, the phrase "especially heinous, atrocious, or cruel, manifesting exceptional depravity" means a conscienceless, or pitiless crime which is unnecessarily torturous to the victim. (15) The defendant intentionally killed the victim while lying in wait. (16) The victim was intentionally killed because of his race, color, religion, nationality or country of origin. (17) The murder was committed while the defendant was engaged in or was an accomplice in the commission of, attempted commission of, or the immediate flight after committing or attempting to commit the following felonies: (ix) Trainwrecking in violation of Section 219. The murder was intentional and involved the infliction of torture. For the purpose of this section, torture requires proof of the infliction of extreme physical pain no matter how long its duration.
(18) The defendant intentionally killed the victim by the administration of poison.
An initiative measure effective June 6, 1990, added two more special circumstances to § 190.2(a)(17): (x) Mayhem in violation of Section 203.
(y) Rape by instrument in violation of Section 289.

Appendix 2
Descriptive statistics for the dependent and control variables used in the multivariate analysis of special circumstances present criminal homicide cases (n = 996).

Appendix 3
Descriptive statistics for number of special circumstances charged by prosecutors in criminal homicide cases with at least one special circumstance present (n = 447).