Predictors of homelessness among vulnerably housed adults in 3 Canadian cities: a prospective cohort study

Homeless and vulnerably housed persons aged 18 or older who were single (i.e. not living with a partner or dependent child) were recruited in Ottawa, Toronto, and Vancouver from January to December 2009. Homelessness was defined as living within the last seven days at a shelter, public space, vehicle, abandoned building, or someone else’s home, and not having a home of one’s own. Vulnerably housed was defined as currently living in one’s own room or apartment, but having been homeless or had two or more moves in the past twelve months. Full-time students and individuals who were visiting the city for three months or less were excluded.

The sampling procedure for recruiting homeless participants was based on the design suggested by Ardilly and Le Blanc (2001) [22]. Study participants were recruited at homeless shelters and meal programs. Homeless participants who did not use shelters were recruited at meal programs proportionally to the estimated number of homeless persons that slept on the street in each respective city. Vulnerably housed participants were recruited from randomly selected rooming houses in Ottawa and Toronto, and from Single Room Occupancy (SRO) hotels in Vancouver. Due to difficulties in gaining access to some of these locations, the recruitment strategy for vulnerably housed individuals was modified to include meal programs, drop-in centers, and community health centers. Data were collected from participants between January 2009 and February 2013. All study participants provided written informed consent and received $20 CDN upon interview completion. The Research Ethics Boards at the University of Ottawa; St. Michael’s Hospital, Toronto; and the University of British Columbia, Vancouver approved this study.

Full details of all survey instruments used in the study have been reported elsewhere [19]. Data on demographic characteristics, health conditions and health status, alcohol and drug use, housing history and quality, social support, legal incidents, and victimization were collected using structured, in-person interviews conducted by trained research personnel immediately following recruitment. Interviews took approximately 60 to 90 minutes to complete. Participants reported their ethnic background based on categories adapted from the Statistics Canada Ethnic Diversity Survey [23].

Chronic health conditions listed in the survey tool were adapted from the Canadian Community Health Survey [24], and participants were asked to report any chronic health conditions that had lasted or were expected to last six months or more and had been diagnosed by a healthcare professional. Lifetime prevalence of mental health diagnoses was determined by self-report. Lifetime prevalence of traumatic brain injury (TBI) was determined using a question from a previous study on prison inmates [25]. Participants were asked whether they had ever had “an injury to the head which knocked you out or at least left you dazed, confused, or disoriented?” Health status was determined using the 12-item Short Form Health Survey (SF-12) to generate Physical Component Summary (PCS) and Mental Component Summary (MCS) subscale scores [26].

Alcohol use was assessed using the Alcohol Use Disorders Identification Test (AUDIT), with a score of eight or more resulting in a positive screen, with scores of 8–15 indicating hazardous, 16–19 harmful, and 20–40 indicating high levels of risk related to alcohol [27]. Drug use problems experienced by participants were assessed using the 10-item version of the Drug Abuse Screening Test (DAST-10) [28]. Scores of three or higher on the DAST-10 resulted in a positive screen, with higher scores indicating moderate (scores 3–5), substantial (scores 6–8), or severe (scores 9–10) drug use problems. The Housing Quality Score was used to determine self-reported quality of the current living environment in 6 domains: comfort, safety, spaciousness, privacy, friendliness, and overall quality [29]. Each item was ranked on a 7-point Likert scale with a maximum total score of 42. Social support was assessed using the Social Support Network Inventory (SSNI), a questionnaire that measured the size of a person’s social network and perceived social support [30].

Housing history data were categorized based on methods adapted from Tsemberis et al. [31]. Each residence in a participant’s housing history was classified into one of 25 types of residence, which were then classified into one of three mutually exclusive residence categories: housed, institution, and homeless. Periods of time spent in institutions were considered periods of being homeless or housed based on a functional classification [31]. Further details are available from the authors upon request.

Participants provided contact information during administration of the baseline survey so that they could be located for follow-up surveys administered approximately one year, two years, and three years after the baseline survey. The follow-up survey included questions of a similar nature to the baseline survey on health status and health conditions, alcohol and drug use, housing status and quality, and social support.

Vulnerably housed participants originally recruited into the study who did not complete any follow-up interviews were excluded from the analyses. Descriptive statistics were used to summarize all quantitative variables. The percentage of vulnerably housed adults who experienced homelessness anytime over the three-year follow-up period was calculated. The main outcome of interest was whether a vulnerably housed participant ever experienced homelessness during any of the three one-year periods between the baseline and follow-up 1 interviews, the follow-up 1 and follow-up 2 interviews, and the follow-up 2 and follow-up 3 interviews.

Baseline characteristics were summarized using means, standard deviations, medians, interquartile ranges, and proportions, wherever appropriate. Comparisons between vulnerably housed participants who did and did not experience homelessness during the three-year follow-up period were performed for baseline characteristics. P-values were calculated from t-test or Wilcoxon rank-sum test for continuous variables. Chi-square test or Fisher’s exact test were used for categorical variables.

Various demographic, health, and housing variables were assessed for an association with a higher probability of becoming homeless over a three-year follow-up period. These characteristics included fixed covariates (determined at the baseline interview) and time-varying covariates (determined at baseline, follow-up 1, follow-up 2 interviews).

The list of candidate predictors of homelessness was developed based on a literature review and consultation with experts. Characteristics assessed for association with experiencing homelessness in the follow-up period were city, time interval, and 1) fixed predictors including: age, gender, ethnicity, highest level of education, percentage of time spent homeless two years prior to baseline divided by 10 for ease of interpretation, number of chronic health conditions (≥3 versus <3), history of a mental health diagnosis, history of TBI, and 2) time-varying predictors evaluated at the beginning of each one-year interval including: employment in the past 12 months, total income in the past 12 months, SF-12 PCS, SF-12 MCS, AUDIT risk level, DAST risk level, housing quality score, and social support network size. For example, SF-12 PCS at baseline, follow-up 1, and follow-up 2 interviews was a time-varying predictor for the main outcome of homelessness during the periods between the baseline and follow-up 1 interviews, the follow-up 1 and follow-up 2 interviews, and the follow-up 2 and follow-up 3 interviews.

Generalized estimating equations (GEE) with the logit link were used to determine the association between predictors and experiencing homelessness, accounting for the correlations between repeated measurements (SAS PROC GENMOD). For fixed predictors, the quasi-likelihood information criteria (QIC) were used to find the correlation structure. For time-varying predictors, we applied the Rotnitzky and Jewell approach [32, 33], and chose the correlation structure for which its associated empirical covariance matrix was closer to the model-based covariance matrix. The GEE model was developed in two steps. Step 1 included city, time interval and fixed predictors, which were retained in the model if significantly associated with or clinically relevant for the outcome (Core Model); Step 2 added time-varying predictors to the Core Model, one at a time. Time-varying predictors significantly associated with or clinically relevant for the outcome were retained in the final model. Analyses for the 2-step process were performed using the exchangeable working correlation structure and coding of time as a continuous variable (time interval years 1, 2, 3) because these settings yielded slightly better goodness of fit statistics. All statistical tests were two-tailed and statistical significance was set at a P-value of 0.05 or less. SAS 9.4 (SAS Institute, Inc., Cary, NC) was used for all analyses.