Eleven‐Month Arrest Outcomes Among Three Crisis Response Models in Michigan

Outcomes related to three crisis response models were compared to law enforcement‐only responses across five sites in the state of Michigan. The three crisis response models were mobile crisis, co‐response models, and office‐based crisis interventions. Sites were defined by the jurisdiction of the crisis response team. If a site did not have a mobile crisis or co‐response team, researchers assessed outcomes of mental health office‐based crisis interventions. At the time of the study, all sites had office‐based crisis interventions, the mobile crisis sites did not have co‐response models, the co‐response sites did not have mobile crisis models, and the office‐based site did not have either a mobile crisis or a co‐response model.

One site offered office‐based crisis interventions: Integrated Services of Kalamazoo is a CCBHC and CMH agency in Kalamazoo County (465 pop/sq mile, 80% White, $70k median household income). Office‐based crisis interventions were predominately performed by clinicians who had existing relationships with the clients; incoming referrals from external agencies (e.g., 911 and law enforcement) were rare. The interventions involved screening for hospitalization, creating safety plans, and scheduling appointments for ongoing care from internal staff referrals. At the time of the study, Integrated Services of Kalamazoo did not have a 24/7 crisis facility, and crisis interventions predominately occurred during daytime hours of 8 am–5 pm Monday to Friday. Integrated Services of Kalamazoo had offered crisis services as a CMH agency since 1964 and became a CCBHC that served a broader client base in October 2021.

Two sites offered mobile crisis interventions: Washtenaw County CMH operated mobile crisis services in Washtenaw County (527 pop/sq mile, 74% White, $87k median household income). LifeWays CMH operated mobile crisis services in Jackson County (224 pop/sq mile, 87% White, $62k median household income). Both mobile crisis teams offered de‐escalation, service linkage, safety planning, and coordination of ongoing care. Both mobile crisis teams responded 24/7 to immediate and follow‐up requests from homes, schools, law enforcement, and other community members. Roughly half of the mobile crisis referrals came from law enforcement, and roughly half came from internal referrals from the CMH agency. Washtenaw and Jackson both offered informal crisis services from the founding of their CMH agencies in the 1960s. Both organizations formally expanded mobile crisis services in 2018 after their counties simultaneously passed millages to support mental health services with property tax dollars. Washtenaw had over 25 crisis staff and Jackson had over 10 crisis staff at the time of the study.

Two sites offered co‐response: Oakland Community Health Network operated a co‐response program with three police departments in Oakland County (Auburn Hills, Bloomfield Township, and Birmingham; 1.5–4k pop/sq mile, 58%–85% White, 77k–$151k median household income). Hegira Health operated a co‐response unit with the police department in the City of Livonia (2.6k pop/sq mile, 84% White, $96k median household income). All four city governments approved municipal funding for the co‐response program. Both co‐response programs received referrals on incidents involving a mental health nexus from law enforcement and dispatch personnel. The co‐response programs responded to “live” incidents or performed follow‐up services on prior incidents. In both sites, a single clinician would perform services alongside law enforcement, rode in a separate vehicle, and offered some follow‐up services without law enforcement presence. Both co‐response programs began in August 2021 and operated on 10‐h shifts, four days per week.

Incident inclusion criteria was defined as a clinical response to a crisis in the year of 2021, outside of an emergency department or jail, to individuals over the age of 18. Office‐based and mobile crisis interventions were operationalized by one of two Medicaid billing codes: H2011 (crisis intervention) and T1023 (pre‐admission screening). Co‐response services were identified through .xlsx spreadsheets maintained by the co‐responding clinicians. Researchers collected information on every intervention that matched the inclusion and exclusion criteria for all sites except Washtenaw, where researchers randomly selected equal samples of calls referred by law enforcement and non‐law enforcement sources.

The crisis service agencies provided researchers access to electronic medical records or .xlsx spreadsheets that contained incident data. If an individual received multiple crisis services in the target year 2021, the first chronological incident was selected as the target crisis event. Researchers documented identifiable data (name, date of birth) and, where available, demographic information (race, sex) for database matching. Researchers read incident narratives to determine the urgency of the response (live or follow‐up), immediate disposition (resolution on scene, call to 911, arrest), referral source (self‐referral, law enforcement, other third party) and incident location type (person’s home, mental health facility, or other public location).

For each of the five sites, researchers requested up to 200 law enforcement reports of mental health crises in the year of 2021. Sample sizes of 200 represented liberal estimates of comparable crisis response cases that the researchers expected to receive, while sufficiently supporting multivariate analyses. One site submitted an additional 200 reports due to a communication error. If the law enforcement agencies maintained a “mental health” or “suicide” category of calls, researchers prioritized them for its comparison group. If the law enforcement agency did not have appropriate categories, the researchers worked with the local dispatch agency to comprise a list of locally‐specific search terms to identify relevant mental health calls via computer assisted dispatch narratives. The dispatch agencies submitted .xlsx sheets containing incidents associated with the search terms, from which the researchers randomly selected n = 200 using a random number generator. Selection criteria and counts of mental health crises across the law enforcement agencies are documented in Table 1. All law enforcement reports included name, date of birth, sex, incident date, and narrative. Research assistants read all police reports to exclude cases that did not involve an obvious mental health component or involved threats to others. Some incidents required discussion between researchers on whether a case included an obvious mental health component, whereupon all coders convened to discuss exclusion criteria until 100% consensus was reached.

Researchers used the Michigan State Police database to determine any arrest in the state. Researchers created a dataset of people who either received a crisis response service or a law enforcement‐only response to a mental health crisis. The submitted data set was matched on names, dates of birth, race, and sex. The Michigan State Police data contained arrest dates, charge level (felony, misdemeanor, ordinance), and charge description. Arrest data was collected for the entire calendar years of 2020 to 2022 to capture a full year before and after each crisis incident in the target year of 2021. To remove arrests related to the target crisis, researchers defined the post‐crisis period as 11 months following the 30 days after the target crisis.

To account for population differences between the law enforcement and crisis response groups, researchers used a quasi‐experimental design with inverse probability of treatment weighting (IPTW). IPTW is a statistical method meant to account for baseline differences that could confound the results of a treatment outcome comparison. By assigning weights based on the inverse of the probability of receiving the observed “treatment” (i.e., receiving a crisis model response), IPTW creates a pseudo‐sample for analysis in which the distribution of covariates is balanced across treatment groups, mimicking the conditions of a randomized trial. Propensity score weighting has been used in prior public health studies to control for population baseline differences (36). Weighted characteristics were age, sex, jurisdiction, and any prior year arrest. Race was a covariate of interest but was missing in 10.5% of cases, systematically related to a response program X²(51) = 207.329, p < 0.001. Researchers excluded race to prevent bias of missingness not‐at‐random (see Table 2).

After weighting, researchers completed univariate (frequencies) and bivariate analyses (Chi‐Square and Mann–Whitney U tests) to describe the study sample, assess the balance between weighted groups, and identify factors of interest that could affect post‐crisis arrest outcomes. Since a pre‐crisis arrest could increase the likelihood of a post‐crisis arrest, researchers also assessed arrest outcomes among the treatment and control subgroups without pre‐period arrests.

Researchers then ran a weighted Poisson regression model to examine the effect of each crisis response model on post‐crisis arrests compared with law enforcement‐only responses. Poisson regression provided an appropriate means for modeling the distribution of non‐negative, integer‐valued outcomes. Researchers hypothesized that controlling for observed characteristics, crisis response models would predict lower incidences of post‐year arrest compared to law enforcement‐only responses. Researchers estimated incidence rate ratios to assess the impact of each crisis response model on the incidence of post‐crisis arrest compared to law enforcement‐only responses.

The unweighted sample was, on average, 40 years of age and equal parts male and female (53%; 47%). The site with the greatest representation of crises was Kalamazoo (25%). While the majority was identified as white (66%), race was unidentified for 10.5% of the sample. There were few violent‐assaultive (4%) or nonviolent/non‐assaultive (11%) arrests in the pre‐period.

A Mann–Whitney U test examining post‐crisis arrest outcomes by crisis response group resulted in statistical significance (p = 0.034); the law enforcement‐only group had on average 0.20 (SE = 0.02) arrests per person in the post‐period compared to 0.11 (SE = 0.02) in the crisis response group, U = 169,984, Z = 2.118, p = 0.034. When comparing subgroups with no pre‐period arrests, the difference was consistent but smaller (M = 0.09, SE = 0.02 vs. M = 0.14, SE = 0.02), while approaching statistical significance, p = 0.092. See Table 3 for the subgroup comparison.

Researchers used a variance inflation factor score tolerance of 5 to diagnose multicollinearity and a factor of 2 as the difference between unconditional mean and variance of error to diagnose overdispersion (37). Finding no evidence of multicollinearity (all variance inflation factor scores under 5) nor of overdispersion (unconditional mean = 0.16 and variance of error = 0.309), researchers identified a Poisson regression model as being appropriate to predict the incidence rate of arrest in the year following the target mental crisis based on the type of crisis response received, controlling by IPTW for age, sex, jurisdiction, and arrest history (37). The model was significant (deviance = 536.66, df = 902, p < 0.001), and the overall fit was good compared to the intercept‐only model (Akaike information criterion = 764.20, Bayesian information criterion = 797.88).

The exponentiated coefficients for all three crisis response models were less than one, ranging from 0.917 (co‐response) to 0.548 (mobile crisis). Only the mobile crisis recipients had a statistically significant lower incidence rate of arrest, 45.2% below the law enforcement‐only group (p = 0.024). Neither the co‐response nor office‐based groups had significantly lower incident rates of arrest compared to the law enforcement‐only group. The incidence rate of arrest among an aggregated crisis response group was not significantly lower than the law enforcement‐only group either. See Table 4 for a summary of weighted Poisson regression results.

While this study controlled for some individual demographics and recent arrest history, groups were nonrandom, and researchers did not have reliable information on crisis severity. Unobserved group differences may have influenced the results. Law enforcement comparison groups are common in the co‐response literature (23, 24), but not in the mobile crisis literature (16, 22). Medicaid billing codes identified the mobile and office‐based crisis services, and law enforcement reports comprised the comparison group. The crisis response model recipients were more likely to have been enrolled in Medicaid than the law enforcement‐only group, suggesting other socioeconomic differences. The office‐based crisis interventions were likely existing CCBHC clients, and the law enforcement‐only group were 911 callers. Law enforcement encounter people experiencing mental health crises who are older, intoxicated, more often diagnosed with mood or psychotic disorders, and violent (22), which may place them at a higher criminogenic risk. Incidents that rise to the level of requiring formal police reports, which contained essential identifiable information for database‐matching, may have been more escalated than typical law enforcement responses to people with mental health illness. This study removed mental health crises that involved threats to others and controlled for arrest history, sex, and age, but could not control for race, substance use or mental illness diagnosis. This study did not have data to determine whether clients connected to treatment, let alone treatment that reduces criminogenic risk. Arrest data did not contain out‐of‐state arrests or other criminal justice involvement, such as probation or parole violations. Arrest patterns in Michigan may not reflect those of other states. Generalizability is limited to areas that had municipal funding to support mobile crisis and co‐response models.

Arrest patterns in the pre‐period years of 2020–2021 were impacted by the COVID‐19 pandemic, and arrest history may be less of a reliable indicator than usual. Researchers did not have pre‐2020 arrest data to account for pre‐pandemic arrests among the cohorts. Compared to aggregate arrests across the studied sites in 2019 (n = 11,280), law enforcement made 21% fewer arrests in 2020 (n = 8901) (40); arrests have since increased, but 2023 arrests remain 11% below 2019 levels (40). Pre‐pandemic arrests would have been useful in the IPTW, but 2019 arrests may overcount contemporary arrest patterns as much as 2020 undercounts them. The co‐response programs were early in their implementation at the time of the target crisis year of 2021, and this study may only reflect implementation growing pains.