Police often serve as first responders to emergency calls involving nonviolent individuals in mental health distress or suffering from alcohol or drug abuse. This procedural norm has been the subject of debate and criticism for two broad reasons. One is that serving as first responders to calls involving mental health crises is a substantial drain on scarce police resources and comes with heavy human and social costs, even in the absence of police violence and use of lethal force ( 1 4 ). Police currently spend more time responding to such “low-priority” calls than to any other type of emergency call ( 5 ). Recent estimates ( 6 7 ) suggest that a quarter to two-thirds of the emergency calls involving disorder, mental health, medical, and noncriminal calls to which police currently respond could instead be directed to mental health crisis experts and other first responders (i.e., a “community response” model). Those charged with minor offenses such as loitering, making false statements, and vandalism cost the criminal justice system roughly $500 to $600 per offense and come with even higher additional social costs ( 8 ). The potential reallocation of resources away from a police response and toward mental health supports is often a part of current initiatives to “defund the police” ( 7 10 ).

Second, having armed and uniformed police as first responders to a mental health or substance abuse crisis may increase the likelihood of costly outcomes and inappropriate care. Individuals living with serious mental illness are no more prone to violence or unpredictability than the general population ( 11 12 ). However, having police officers as first responders to a mental health crisis can result in unnecessarily violent and tragic outcomes ( 13 14 ). Recent news coverage ( 15 17 ) has drawn public attention to particularly shocking incidents in which responding police officers seriously harmed or killed a person in mental health distress ( 18 ). More generally, having the police respond to such incidents can be costly and unproductive because police are more likely than mental health clinicians to direct individuals experiencing a mental health episode to the criminal justice system rather than to the appropriate health care ( 13 ).

In response to these concerns, municipalities across the country have begun to pilot targeted reforms. The two most common approaches augment the capacity of police officers to serve as effective first responders to individuals experiencing mental health crises. The “crisis intervention team” (CIT) approach emphasizes training police officers how to respond to individuals in crisis and connect them with appropriate services ( 19 ). In contrast, the “co-response” model involves structuring explicit partnerships between police departments and professional mental health practitioners so they can simultaneously respond to incidents involving mental health crises ( 20 23 ). A third and less common approach either delays or foregoes on-scene police involvement in certain incidents by relying on “a new branch of civilian first responders known as ‘Community Responders’” ( 5 ). These so-called community response programs can use first responders with expertise in a breadth of social service support and establish a triage protocol under which emergency calls for mental health crises are first addressed by a health team (e.g., a mental health crisis interventionist and a paramedic) before deciding whether to request direct police involvement ( 5 ).

The momentum behind the adoption of programs that seek to improve police interactions with individuals in mental health crises has motivated multiple empirical studies that seek to understand their impact. Several systematic reviews and meta-analyses have synthesized this evidence, particularly focusing on the more common CIT and co-response models ( 20 28 ). In general, this empirical literature suggests that these program innovations have beneficial effects by reducing arrests and detention rates, but evidence is mixed on whether these programs are cost-effective. However, the research designs used in these studies (e.g., case notes, qualitative and descriptive studies, before/after comparisons, and cross-sectional comparisons) generally do not support credible causal inference. For example, one recent prominent review concludes that “… we caution against drawing conclusions related to causality based on these findings” ( 27 ). There is a similar lack of evidence on the impact of less common community response models. Existing evaluations are typically conducted internally by cities, police departments, or community response teams and rely on descriptive evidence of the number of calls taken by the few community response units operating across the United States ( 5 ).

Furthermore, critics warn that initiatives to reduce police involvement in response to emergency calls will “embolden the bad guys” ( 29 ) and unintentionally increase the prevalence of more serious criminal offenses ( 30 33 ). This belief, often referred to as the “broken windows” theory in which police response to low-priority criminal violations prevents more serious ones, underscores the need for research studies that can provide credibly causal estimates of the impact of these innovative programs both on the focal, less serious crimes they target and on more serious offenses. However, the debate on defunding police has a limited causal basis, having “proceeded without adequate research about either the scale or nature of issues that the police handle or the potential consequences of the proposed reform efforts” ( 7 ). This study seeks to provide such evidence by examining the impact of a community response program recently piloted in the City and County of Denver, Colorado through the independent analysis of a preregistered, quasi-experimental design coupled with several complementary robustness checks.

Operators responding to 911 calls for assistance dispatched STAR staff to eligible incidents that were located in the designated police precincts and during the program’s hours of operation (Monday to Friday, 10 a.m. to 6 p.m.). The identification of emergency calls eligible for STAR services relied on two specific screening criteria. First, the incident had to designate at least one of several codes: calls for assistance, intoxication, suicidal series, welfare checks, indecent exposure, trespass of an unwanted person, and syringe disposal ( 40 ). Second, to dispatch the STAR van, there needed to be no evidence that the incident involved serious criminal activity, such as weapons, threats, or violence, or serious medical needs. The STAR team also responded to calls from uniformed police to engage with community members in crisis and initiated engagement in the field on their own. Over the 6-month pilot period, the STAR team responded to 748 incidents or nearly 6 incidents per 8-hour shift. Roughly a third of calls to STAR occurred at the request of responding police, while the rest were due to a direct 911 dispatch or to the STAR team responding independently to a field observation—none of which required a call to police for assistance or for a response to a criminal offense ( 41 ).

STAR began operations on 1 June 2020 for a designated 6-month pilot period. During this period, STAR limited its operations to selected 911 calls for assistance in eight purposefully chosen police precincts (i.e., out of the city’s 36 precincts), where the need for STAR services was anticipated to be the greatest. The pilot area was in the central downtown area of Denver (fig. S1) and largely represents neighborhoods with residents who are more affluent, educated, and white than the city as a whole (see table S1). However, all but one of the neighborhoods in the STAR pilot service area are also designated by the city as “displacement-vulnerable” areas, rapidly gentrifying city spaces where poor and otherwise at-risk residents are being pushed out ( 35 ). In such contested urban spaces, there are often increasing demands on police to conduct “rabble management” that addresses overwhelmingly nonviolent incidents ( 36 39 ).

The Support Team Assistance Response (STAR) program in Denver provides a mobile crisis response for community members experiencing problems related to mental health, depression, poverty, homelessness, and/or substance abuse issues. The STAR response consists of two health care staff (i.e., a mental health clinician and a paramedic in a specially equipped van) who provide rapid, on-site support to individuals in crisis and direct them to further appropriate care including requesting police involvement, if necessary. The design of the STAR program is based on the Crisis Assistance Helping Out On The Streets program developed in Eugene, Oregon ( 34 ).

Measuring STAR impacts on crime

We identify the impact of the STAR program on STAR-related and STAR-unrelated measured crime using “difference in differences” (DD) and “difference in difference in differences” (DDD) designs that effectively rely on before-after comparisons across treated and comparison precincts (i.e., along with the evidence from several complementary robustness checks and alternative estimation procedures). To identify the impact of the STAR program, we consider all criminal offenses reported by the City and County of Denver through data collected as part of their participation in the federal National Incident-Based Reporting System (NIBRS). These data include calls for police assistance that escalated to offenses reported by the police regardless of whether they led to formal charges (including arrest) or whether the STAR team was dispatched or responded to the call. However, offenses are not differentiated by whether they led to an arrest or some other offense-related outcome (e.g., a citation). We also do not have access to such arrest and citation data and recognize their possible confounding. That is, such data may be missing (e.g., valid criminal offenses where no one is apprehended and, therefore, there is no arrest or citation) and may be confounded further by police and prosecutorial discretion around whether to sustain an arrest or citation for a given offense.

Before our analysis, we coded each offense as directly related to STAR operations (e.g., disorderly conduct, trespassing, alcohol, and drug use) or not (e.g., burglary; see tables S2 and S3 and Supplementary Text for details). For this focal outcome (i.e., lower-level reports of criminal offenses), we expect either the STAR team or the police to often engage the individual in question. If a criminal offense is recorded during such service calls, then it implies either an arrest or a citation. If no crime is recorded, then it implies either a field determination that no criminal offense occurred or a discretionary decision not to record such low-level criminal offenses (e.g., trespassing).

The impact of the STAR program on the frequency of these offenses is theoretically uncertain. For example, to the extent police who respond to mental health and substance abuse incidents consistently direct individuals in crisis to health care services without also identifying them as low-level criminal offenders, the overall effects of the STAR program would be muted—or even null. The Denver police have participated in CIT training designed to support their capacity to identify individuals who need mental health support and to direct those individuals to appropriate care. Because the comparison condition in this study consists of such CIT-trained police as first responders, the introduction of the STAR team could, in theory, have small or nonexistent effects on recorded crime.

This study provides quasi-experimental evidence on the overall (i.e., “reduced form”) effect of STAR’s community response approach on the number of recorded crimes. With respect to reducing recorded criminal offenses, the STAR program’s overall impact could reflect the combination of two broad, underlying mediating mechanisms that merit careful emphasis. One involves program-induced reductions in the recording of existing criminal offenses, while the other concerns reductions in actual crime.

The first mediating mechanism would occur when STAR first responders simply do not record an existing criminal offense (e.g., substance abuse and disorderly conduct) that police officers would record when responding to a given incident. This reporting mechanism has clear empirical relevance given that (i) under NIBRS procedures, law enforcement officials (i.e., not STAR staff) identify and report criminal offenses and (ii) the program data from the period we study indicate that STAR staff did not involve police in their service calls ( 41 ). However, this reporting mechanism also reflects an impact of social consequence. Specifically, it implies that, when STAR staff replace police as first responders, individuals in mental health or substance abuse crises may be more likely to receive health care and are less likely to be identified as criminal offenders (i.e., implying arrests or citations).

A second class of mediating mechanisms underlying STAR’s overall impact also reflects effects of clear policy relevance. Specifically, there are several reasons that the STAR program could also lead to a genuine reduction in the prevalence of criminal offenses. First, this would occur if the STAR team is more effective than police in implementing de-escalation tactics that reduce the likelihood of further criminal acts (e.g., assaults) when responding to an incident ( 42 43 ). In addition, the STAR team may prevent crime in the near future by reducing recidivism among individuals in crisis. Individuals experiencing mental health or substance abuse crises are quite likely to reoffend ( 42 ). However, STAR’s targeted provision of health care could reduce the prevalence of such future incidents that would otherwise be recorded as crimes.

In addition, there are at least two other potential “spillover” mechanisms by which the STAR program reduces crime. One is the possibility that the presence of the STAR program in a precinct improves police officers’ implementation of their CIT training. This can occur if officers are more likely to call the STAR team when in need or if they better implement their own CIT training by independently directing individuals in mental health crises to health care responders rather than the criminal justice system when they know STAR is active in their precinct. Another possible, although uncommon, mechanism happens when STAR staff initiate a response in the field. If STAR staff happened to observe an individual clearly in need of their services, then they would sometimes respond without evidence that a crime had yet occurred and without direction from 911 dispatchers or on-scene police officers ( 41 ).

A reduced-form analysis similar to the one used in this study cannot exactly decompose STAR’s overall impact on crime into the components attributable to these varied mechanisms (e.g., genuine crime reductions and the differential recording of individuals in mental health crises as not having committed criminal offenses). However, we do discuss two pieces of ancillary evidence that indicate whether STAR’s overall impact partly reflects lower levels of actual crime in addition to the reduced reporting of criminal offenses. First, we provide direct evidence for genuine crime reductions by examining the impact of the STAR program on crimes occurring outside STAR’s operating hours (e.g., spillover benefits due to reduced recidivism). Second, we compare the total crime reduction attributable to STAR to the amount that would be expected if STAR’s effects only operated through its service calls. We construct this expected number by multiplying the number of STAR service calls conducted during the pilot period by the number of criminal offenses typically recorded in such criminal incidents during the pretreatment period.