Study design and setting
We conducted a two-year randomized controlled trial in Toronto, part of the At Home/Chez Soi cross-site project, in order to assess the effectiveness of HF with intensive case management for homeless adults with mental illness from diverse ethnic minority groups. The protocol for the overall At Home/Chez Soi project  and the Toronto site specifically  have been described previously. Prior to randomization, all participants were first stratified either to a high needs or moderate needs group based on their level of need for mental health services. At the Toronto site, participants in the moderate need group were further stratified by ethnicity (belonging to ethnic minority groups or not). This article focuses exclusively on study outcomes among the sample of moderate needs participants who belong to an ethnic minority group.
The study was approved by the Research Ethics Board (#09-208) of St. Michael’s Hospital in Toronto, and was registered with the International Standard Randomized Control Trial Number (ISRCTN 42520374). All study participants provided written informed consent.
All At Home/Chez Soi participants met three inclusion criteria: a) ≥ 18 years old; b) absolutely homeless (having no fixed place to stay for at least the past seven nights with little likelihood of finding a place in the upcoming month) or precariously housed (currently occupying a single room in a multi-tenant building or house with shared common areas including bathroom and kitchen  or a hotel/motel as a primary residence, and having a history of two or more episodes of absolute homelessness in the past year); and c) presence of a current mental disorder with or without a co-existing substance use disorder, based on the DSM-IV  criteria in the Mini International Neuropsychiatric Interview 6.0 (MINI) . Eligible diagnoses included: a) major depressive episode; b) manic or hypomanic episode; c) mood disorder with psychotic features; d) panic disorder; e) posttraumatic stress disorder; or f) psychotic disorder. Diagnosis of a substance use disorder alone did not qualify participants for study entry. Participants were excluded from the study if they: a) were relatively homeless (people who are residing in conditions that do not meet basic standards but who are not absolutely homeless or precariously housed, including those living in overcrowded or hazardous housing, transitional housing such as shelters for domestic abuse, long-term institutions, couch surfing, and people at risk of homelessness or lacking a dwelling for a short period of time due to disasters such as fire or economic situations) ; b) lacked a diagnosis of a serious mental disorder; c) had no legal status in Canada; or d) were current recipients of mental health supports via assertive community treatment (ACT) or intensive case management (ICM) services .
Participant recruitment was based on referrals from both community social service agencies and acute care services. A targeted recruitment strategy was employed to ensure that the broader study sample adequately represented the target population , and all participants were assessed by the intake coordinator who determined study eligibility. Given this study’s focus on black and ethnic minority participants, we recruited extensively among diverse ethnic minority groups through outreach to ethnoracial agencies serving this population.
Stratification and randomization
Prior to randomization, all eligible participants for the larger trial were stratified into either a high need or moderate need group, indicating their need for mental health services. The stratification algorithm defined high need participants as those who had: 1) community functioning scores from the Multnomah Community Ability Scale (MCAS) ≤ 62 ; and 2) a diagnosis of current psychotic or bipolar disorder based on the MINI , in addition to meeting one of the following three criteria: i) ≥ 2 hospitalizations for mental illness in any one year in the last 5 years; or ii) diagnosis of comorbid substance use based on the MINI; or iii) recent arrest(s) or incarceration(s) in the past 6 months. High needs participants were randomized to HF with ACT or a usual care group. All other participants were considered to have moderate need for mental health supports services.
Moderate needs participants were further stratified by ethnicity. Self-reported ethnicity was measured using a form validated locally , which asked people to select their ethnicity from one of 15 groups based on race and geographical origin: those who selected Black African (e.g. Ghana, Kenya, Somalia), Black Canadian/American, Black Caribbean (e.g. Jamaica, Trinidad, Tobago), East Asian (e.g. China, Japan, Korea), Indian-Caribbean (e.g. Guyana with origins in India), Latin American (e.g. Argentina, Chile, Costa Rica), Middle Eastern (e.g. Egypt, Iran, Israel, Palestine), South Asian (e.g. India, Pakistan, Sri Lanka) and South East Asian (e.g. Malaysia, Philippines, Vietnam) or who reported mixed background that included at least one of the ethnic groups listed above were considered eligible for the adapted HF program. Participants were excluded if they self-identified as Aboriginal, White (European or Canadian) or of a mixed ethnicity that did not include one of the specified groups listed above.
Moderate need participants not belonging to an ethnic minority group were allocated to a regular HF with ICM program or usual care control group. All moderate needs participants belonging to an ethnic minority group (n = 237) were randomized to the adapted HF intervention or usual care. A small group of participants (n = 33) randomized to the adapted HF program did not receive the adapted intervention either because there was no space available or they requested a non-ethno-racial focused program, and instead received services from the regular HF with ICM program.
Randomization took place via adaptive randomization procedures using a laptop computer connected to the study data management centre: by continuously adjusting the probability of allocation to each treatment group based on existing group assignment, this procedure can produce better balance between treatment groups than strict randomization in small and moderate sized studies [10, 19]. Several aspects of the study prohibited blinding, including the nature of the administered questionnaires (detailed housing history and service use), location of participant interviews (some participants elected to be interviewed at their place of residence) and follow-up procedures (locating participants often required aid from case managers or community workers, where consent was given). As a result, masking follow-up data was also not possible. However, the allocation algorithm was concealed from both the participants and the research staff. Following randomization, participants allocated to the intervention group were immediately connected with their treatment teams, while usual care participants and their referral sources were provided with information about other existing services.
The adapted HF intervention was developed uniquely for the Toronto site with ICM services provided by a mental health agency exclusively serving ethnic minority groups using anti-racist and anti-oppressive frameworks of practice. The model has been described in greater detail elsewhere . The development of the adapted model was informed by practices within the leading agency and a review of the literature [11, 20]. Participants were provided with immediate access to permanent housing of their choice in their preferred neighbourhood (via rent supplements of $600 CAD paid directly to the landlord), in addition to individualized and client-driven mental health support services. Participants worked with a case manager to develop a participant-driven treatment plan, which included both immediate and long-term goals, such as application for disability benefits, access to primary care or other health services in the community, reconnecting with social networks, participation in substance misuse treatment programs and vocational training . The staff ratio in the adapted HF team was 17:1 and services were provided for the duration of the follow-up.
The main principles of anti-racist/anti-oppressive service delivery have been outlined elsewhere [11, 20] and include: empowerment, education, alliance building, language use, alternative healing strategies, advocacy, social justice/activism and fostering reflexivity (critical self-knowledge, awareness and examination of one’s social position and its influence) . HF and anti-racist/anti-oppressive practices share several elements, including a focus on client empowerment and choice, with HF additionally offering targeted strategies to ensure housing stability . The agency offering the adapted HF intervention was committed to anti-racist/anti-oppressive frameworks of practice across program structures and offered regular staff training in such practices, as well as linguistic and culturally accessible programming and services, and a physical environment that is inclusive and welcoming to people from ethnic minority groups .
Anti-racist and anti-oppressive staff practices focus on breaking the silence about racism, addressing racism and discrimination, and examining power inequities, oppression and mental health together, taking anti-racist and anti-oppressive action as required [11, 20]. Case managers embrace client-centred, strengths-based, holistic approaches to mental health, recognizing the importance of community of origin, family, and different ways of healing [11, 20]. A key element of the adapted HF program is that case managers were representative of the population they served whenever possible and fluent in the primary language of program participants. In 2009, the agency offered services in 18 languages in addition to English. In addition to ICM services, the agency offered art and music therapy, a community kitchen, computer programs, life skills, traditional Chinese medicine, yoga, English as a second language, as well as support groups for men, women and youth [11, 22]. Another key treatment approach adopted by the agency was the inclusion of families and peer networks early in the recovery process [11, 22].
Individuals randomized to the usual care group were able to access a variety of traditional housing programs, mental health and community services available in the city of Toronto and were provided with information on how to access such services. Toronto is a service rich environment with a variety of health and mental health services, as well as programs specifically serving people experiencing homelessness. Local services include crisis programs, drop-in centres, emergency shelters, inpatient/outpatient mental health services, meal programs, street outreach programs, supportive housing programs, comprehensive primary care teams, ACT and ICM teams . In addition, a variety of specialized primary care and community mental health agencies exist which serve specific ethnic or language groups, immigrants and refugees .
Participants were met every 3 months for face-to-face, structured, laptop computer-assisted interviews. Longer interviews took place at baseline, 6-, 12-, 18- and 24-months. During each interview, data was collected and entered wirelessly directly into a secured central database. Brief call-in updates and interviews were conducted with participants on a regular basis to maintain contact and improve follow-up rates. Participants were financially compensated for all interviews and updates ($10 CAD for monthly calls, up to $40 for short interviews and $60 for longer interview). More details on study methodology and design for the At Home/Chez Soi study  and on this adapted program, can be found elsewhere .
The prespecified primary study outcome was housing stability, evaluated using the Residential Time Line Follow-Back (RTLFB) Inventory . For each participant, we calculated the percent of days spent in stable housing during the 24-month follow-up period. Stable housing was defined as living either in one’s own apartment, house or room or with family in addition to either tenancy rights or an expectation of residing at this location for ≥ 6 months, as opposed to unstable housing, which included living on the street, temporary residences (< 6 months’ duration and/or no tenancy rights), shelters, crisis units and institutions.
Secondary outcomes explored mental and physical health, social outcomes and services use. Generic quality of life was assessed using the “overall health” Visual Analogue Scale (VAS) of the EuroQol 5 Dimensions (EQ-5D), which measures self-rated health status (both physical and mental) along a scale that ranges from worst (0) to best (100) imaginable health state . Severity of problems with drugs or alcohol within the past month was assessed using the substance use screener (SDScr) of the GAIN Short Screener, GAIN-SS . Additional questions asked about drug and alcohol use related problems and the amount of money spent on alcohol and drugs in the past-month. Psychiatric symptoms within the past month were evaluated using the 14-item modified Colorado Symptom Index (CSI), for which a total score was calculated . Community functioning was evaluated via the total summary score of the MCAS . The total score of the Quality of Life Index (QoLI-20), an instrument used widely in this population, was used for assessing condition-specific quality of life . Physical and psychological community integration were assessed using the separate physical and psychological subscales of the Community Integration Scale (CIS) . Emergency department use was assessed with two questions from the Health Service and Justice Service Use Questionnaire (HSJSU): i) “In the past 6 months, have you been to a hospital emergency room?” and ii) “Approximately how many emergency room visits did you have in total?” The number of days hospitalized was calculated from data collected from the RTLFB by examining the number of overnight stays in hospitals over the study period. The number of police arrests was captured by two additional questions from the HSJSU: i) “In the past six months, have you been arrested?”, and ii) “How many times?”.
Sample size calculation
The primary outcome of housing stability was used to estimate a clinically meaningful effect size. In a previous randomized trial of supportive housing and ICM compared to standard care among homeless veterans with mental illness and/or substance use , participants spent an average of 66 % and 53 % of days housed (of past 90 days) in the intervention and usual care groups, respectively, with an estimated common group standard deviation as 26 %, resulting in a medium effect size (Cohen’s d = 0.5) . In order to account for an assumed attrition rate of 40 %, we estimated that a sample size of 100 participants would be necessary in each arm in our study to provide 80 % power to detect a medium effect size (d = 0.5)  at 24 months using a two-sided t-test.
Missing data occurred in the key outcomes due to participant withdrawal, non-responses or refusal to answer, which we decided to impute a priori using sequential regression multivariate imputation (SRMI) . This method allows for efficient imputation by fitting a model to each variable, conditional on all others, and imputing one variable at a time . Variables included in the multiple imputation model included a) outcome variables collected at all study visits; b) study site; c) age at enrolment; d) gender; e) ethnic minority status and f) Aboriginal status. Imputed values were restricted to the theoretical range of the original variables by use of bounds. Using this approach, 40 imputed datasets were created using the mi impute chained command in STATA v.13 (StataCorp LP) and results were combined using PROC MIANALYZE in SAS 9.4 (SAS Institute). Because the extent of missing data was small (5 %), no imputation was performed for percent of days stably housed or the number of hospitalizations, which were calculated for the duration of the follow-up period for each participant for whom housing data was available, and who was known to be alive.
All statistical analyses were performed with SAS version 9.4 on an intention-to-treat basis. The primary outcome (percent of days stably housed during the 24-month follow-up period) and the hospitalization outcome (rate of hospitalization during the 24 months follow-up period) were analysed by fitting mixed models, which assessed the main effect of treatment (adapted HF vs. usual care). For the percent of days stably housed, a linear mixed model was used (PROC MIXED), while a zero-inflated negative binomial (PROC GENMOD) model was fitted for the number of days hospitalized. For all other secondary outcomes for which longitudinal data was available, analyses were conducted using linear mixed models (PROC MIXED) for continuous outcomes and generalized linear models (PROC GENMOD) for count variables. The main effects of time and treatment as well as the interaction of treatment*time were examined. The unstructured covariance matrix for repeated measures was considered in all models. The significance level was set at 5 % for all analyses.
In addition to the outcome analyses, we also examined what baseline variables were associated with the duration of being housed during the study period. In these analyses, the housing outcome was dichotomized into: a) the reference group, which was defined as those who spent less than half of the time (< 50 %) in stable housing during the study period; and b) the group of interest, which was defined as those who had been stably housed half or more of the time (≥ 50 %) during the 24-month study period. These secondary analyses employed a multivariate logistic regression analysis, and were conducted using a two-step process:
A list of baseline variables was established, based on potential associations with the outcome variables of interest, and included self-reported demographic variables (age, gender, total length of time homeless, education level, immigrant status, length of residency in Canada, native language), clinical variables (specific MINI diagnoses, including psychosis, major depressive disorder, post-traumatic stress disorder, alcohol or substance abuse or dependence) and the number of self-reported chronic health conditions . Independent logistic models were performed for each baseline variable to test for associations. With the exception of the model examining the treatment variable, all other Step 1 models were adjusted for treatment group (i.e. these were bivariate tests). An p < 0.20 was set for these preliminary Step 1 tests, as using an p < 0.05 to examine potential confounders can lead to deletion of important confounder variables from the model .
All covariates that were identified as significant in Step 1 at p < 0.20 were added into a multivariate logistic regression model. If two or more of the selected variables from Step 1 were highly correlated (e.g. age and length of time in Canada), we performed separate models with each of these variables, but only retrained one of the highly correlated variables in the multivariate model based on strength of effect in the model.