Impact of Mobile Crisis Services on Emergency Department Use Among Youths With Behavioral Health Service Needs

Details about Connecticut’s mobile crisis model of service delivery, implementation, support services, performance measures, and outcomes have been described elsewhere (23). The program provides coverage to youths who are age 18 and younger as well as to older youths who are still attending high school via six primary contractors and 14 total sites located throughout the state. Clients anywhere in the state can access mobile crisis services by dialing a specifically assigned number (211) that connects to a call specialist at a call center. Each site has a clinical team consisting of a site director and multiple master’s-level clinicians (drawn from social work, psychology, marriage and family therapy, and related fields) as well as access to a child and adolescent psychiatrist to provide psychiatric evaluation and medication management.

The team provides crisis stabilization and support; screening and assessment; suicide assessment and prevention; brief, solution-focused interventions; and referral and linkage to ongoing care. An episode of care can last up to 45 days. The service time frame allows for stabilization of the presenting concern and for transition to potential additional psychiatric care and services that may be needed (outpatient services, extended day treatment, evaluation for inpatient hospitalization). Mobile crisis services can be accessed repeatedly by youths and families with no set limitation.

All data pertaining to ED use in this study were based on paid Medicaid claims. Accordingly, youths included in this study had to be Medicaid-eligible and enrolled during the study period. Deidentified data were made available from Beacon Health Options, the behavioral health administrative service organization (ASO) for the state Medicaid agency (the Connecticut Department of Social Services). Because ED service use was an outcome, we collected data on ED visits within a specified time frame, specific information about client characteristics, and information about the specific types of medical services (whether the visit was behavioral health or medically related ED treatment or inpatient psychiatric treatment) and associated ICD-9 psychiatric diagnoses for each ED visit. Accordingly, we used data from the ASO to create a client sample and a comparison group sample.

Data from clients receiving Mobile Crisis services during Connecticut’s fiscal year (FY) 2014 (July 1, 2013, to June 30, 2014) were obtained from the Provider Information Exchange (PIE) system that is maintained by the Connecticut Department of Children and Families (DCF). For these youths, we used diagnoses data from the PIE system, which contains up to 10 ICD-9 diagnostic codes derived from information obtained at intake and discharge. DCF PIE data were matched to Medicaid utilization data for the 18 months before and after the individual’s identified “index” mobile crisis episode (the first mobile crisis episode in FY 2014).

To obtain ED service information for the sample of youths who received mobile crisis services, the ASO used a probabilistic matching procedure between the sample file and Medicaid utilization data. All clients were younger than age 18 at the time of their index episode and had no more than 31 days of Medicaid ineligibility during the 18-month period before or after their identified Mobile Crisis episode. Youths younger than age 3 were excluded from the sample. The final client sample was composed of 2,532 youths. The mean number of days of service per episode for this sample was 15 days; the median was 9 days. Service episodes ranged from 0 days (the mode) to a maximum of 426 days.

All members of the comparison sample had at least one behavioral health ED Medicaid claim (any ICD-9 behavioral health diagnosis in any of the first four positions on the paid claim) during FY 2014 and no mobile crisis service episode during the same time period. Members of the comparison sample were younger than age 18 at the time of the index episode (there were no youths younger than age 3 in the sample), had no more than 31 days of Medicaid ineligibility, and were not dually enrolled (not enrolled in both Medicaid and Medicare) during the 18-month period before or after their ED index episode. An index episode in the comparison sample was defined as the first behavioral health ED visit during FY 2014 (between July 1, 2013, and June 30, 2014). The final comparison sample was composed of 3,961 youths.

ED outcomes were determined by looking at service counts, dates, and diagnoses for behavioral health ED visits and medical ED visits by using Medicaid claims data. In both settings, the ASO counted a behavioral health ED outcome as one in which any psychiatric diagnosis was provided during an ED service encounter. Analyses looked at whether there was any ED service encounter with a psychiatric diagnosis (binary outcome) as well as the number of ED service encounters (visits) with psychiatric diagnoses (continuous count outcome) in the 18-month follow-up period.

Data acquisition and analyses were approved by the University of Connecticut and the Connecticut DCF institutional review boards.

We constructed propensity scores (see the next section) and conducted stratified analyses within propensity score quintiles (24, 25). In one set of analyses, the binary indicator (any post-episode ED visit) was regressed on an indicator of group status by using logistic regression. In another set of analyses, taking into account the distribution of the dependent variable, the count variable (number of post-episode ED visits) was regressed on the same indicator by using negative binomial regression (26). SPSS, version 25 (27), was used to perform the regressions. After regressions were carried out, summary odds ratios (ORs) were constructed for the logistic regression and incidence risk ratios (IRRs) were constructed for the negative binomial regression by using meta-analysis procedures in Stata, version 15 (28).

To create propensity scores, we selected a group of variables that was potentially predictive of mobile crisis service use. These variables were as follows: region where a child resided during his or her index episode (Connecticut is divided into six service regions; region was therefore a five-category variable with one reference group), race-ethnicity (a four-category variable constructed as follows: non-Hispanic White, non-Hispanic Black, Hispanic, and other), gender, age, number of prior behavioral health–related ED visits (in the 18 months before the index episode), and diagnosis. For diagnosis, we used ICD-9 codes available for each group at the time of the index event (the 10 possible data fields of the PIE data for the client sample and from the four possible data fields of the Medicaid claims data for the comparison sample) to recreate non-mutually-exclusive, single-level Clinical Classification Software (29) codes as follows: adjustment disorders; anxiety disorders; attention deficit, conduct, and disruptive behavior disorders; depressive disorders; other (nondepression) mood disorders; and low prevalence diagnoses.

Low prevalence diagnoses were all diagnoses with a prevalence of less than 10% in the client sample. This category included nine Clinical Classification Software categories rarely diagnosed in the client sample (delirium, dementia, and amnestic and other cognitive disorders; developmental disorders; impulse control disorders; personality disorders; schizophrenia or other psychotic disorders; alcohol-related disorders; substance-related disorders; screening and history of psychiatric and substance use codes; miscellaneous psychiatric disorders). Since diagnostic categories were binary variables and not mutually exclusive categories, youths could have more than one diagnosis.

Table 1 shows the distribution of the variables used for construction of propensity scores. The distribution suggests significant differences between the two groups on all of the 11 variables used for propensity score construction. With respect to region, the client sample was overrepresented in regions 1 and 4; the comparison sample was overrepresented in region 3. With respect to race-ethnicity, the client sample had higher proportions of non-Hispanic black youths and Hispanic youths relative to the comparison sample. There were relatively more females than males in the client sample than in the comparison sample. The mean age was slightly younger for the client sample (however, the mean age was around 12 years for both groups). There were significantly higher rates of adjustment disorders, depression, and other mood disorders in the client sample. The comparison sample had significantly higher rates of diagnoses grouped into the low prevalence category. The mean number of behavioral health ED visits prior to the index episode was higher for the client sample than for the comparison sample.

More than one in five youths in the comparison sample had any prior behavioral health ED visits compared with more than two in five youths in the client sample. Thus, in addition to demographic differences, there were clear clinical and service utilization differences between the two groups, with the youths in the client sample showing more extensive prior psychiatric service use and potentially greater psychiatric service treatment need (given the diagnostic differences) and acuity (given prior behavioral health ED utilization).

Client sample status was regressed on the 11 variables noted earlier for the two combined groups (the client sample and the comparison sample). The logistic regression was used to generate propensity scores (the probability of being seen by mobile crisis services, given the combination of covariates) for each case. Youths in each group were then stratified into quintiles on the basis of their propensity score values. Regardless of the service actually received, those in the first quintile presented as least likely and those in the fifth quintile as most likely to receive the mobile crisis service. The frequency distribution for youths in each of the two groups within each quintile is shown in Figure 1.

It is reassuring to note that both groups are represented in all five quintiles, suggesting that valid comparisons are possible. As might be expected, youths in the comparison sample were markedly overrepresented in the first quintile, and youths in the client sample were overrepresented in the fifth quintile. To evaluate the effectiveness of our propensity score construction, we examined whether the differences shown in Table 1 persisted when comparisons were stratified within quintile (not shown here). The only variable that remained significantly different across the two groups within quintiles was an indicator of whether a youth had a behavioral health ED visit prior to the index episode. Accordingly, we retained an indicator of prior behavioral health ED use in subsequent regression analyses conducted within propensity quintiles.

Overall, 45.1% (N=2,930) of the combined sample had an ED visit associated with a psychiatric diagnosis during the follow-up period. Within the client sample, the percentage with any ED visit was 43.5% (N=1,102; range 1–32 visits); within the comparison sample, the percentage was 46.1% (N=1,929; range 1–39 visits). Conversely, 56.5% (N=1,430) of youths in the client sample and 53.9% (N=2,133) of youths in the comparison sample had no ED visits at follow-up. These data are summarized in Figure 2, which underscores the similarity of the two groups with respect to behavioral health ED outcomes before propensity score stratification.

We constructed five separate logistic regression models (one within each quintile). In these models, an indicator of any behavioral health ED use in the 18 months after the index visit was regressed on group status (client sample vs. comparison sample) and a binary indicator of any behavioral health ED use in the 18 months before the index episode. The logistic regression models yielded significant ORs in three of the five quintiles (quintiles 1, 4, and 5), suggesting a trend toward reduction in risk for subsequent behavioral health ED visits for the client sample. The reduction in risk is visually illustrated in Figure 3, where point estimates for ORs (squares with dots) are all to the left of the value of 1 (which is represented by a vertical line passing through each quintile). The largest decrease in risk is in the first quintile (OR=0.53); the ORs are 0.66 and 0.61, respectively, for the fourth and fifth quintiles. The pooled (overall) OR of 0.75 (Figure 3; 95% confidence interval [CI]=0.66–0.84) indicates that overall, youths in the client sample had a significant reduction in odds of a subsequent behavioral health ED visit compared with youths in the comparison sample.

We constructed five separate negative binomial regression models (one within each quintile). This model was chosen on the basis of the distribution of the outcome measure, which, for both groups, had a large number of zero values (the mean and median for both groups was zero) and a wide dispersion of values, with a few extreme values reflecting a large number of behavioral health ED episodes. In these models, a continuous indicator of the number of behavioral health ED visits after the index episode was regressed on group status (client sample vs. comparison sample) and, as in the logistic regression models, a binary indicator of any behavioral health ED use in the 18 months before the index episode. The negative binomial regression models (Figure 4) yielded significant IRRs in quintiles 4 and 5. In quintile 4, the IRR was 0.80 (95% CI=0.65–0.99); in quintile 5, the IRR was 0.64 (95% CI=0.52–0.80). The pooled (overall) IRR value of 0.78 (Figure 4; 95% CI=0.71–0.87) suggests that youths in the client sample had a significant reduction in the incidence of behavioral health ED visits in the 18 months after the index episode compared with youths in the comparison sample. We ran and plotted three additional models to evaluate the robustness of the findings obtained from the negative binomial regression (see Supplemental Figure 1): predicted average probability distribution of behavioral health–related ED service use counts from Poisson, zero-inflated Poisson, and zero-inflated negative binomial models. On the same figure, we also plotted the observed probability distribution of behavioral health–related service use counts. The considerable overlap between the predictions yielded by the negative binomial and the zero-inflated negative binomial models suggests that the zero-inflated negative binomial model does not significantly improve model fit, and supports the use of the more parsimonious negative binomial model.