Housing First, Connection Second: The Impact of Professional Helping Relationships on the Developmental Trajectories of Housing Stability for People Facing Severe and Multiple Disadvantage

Background . This study sought to establish the impact of professional helping relationships on the developmental trajectories over 24 months of housing stability for 2,141 people facing severe and multiple disadvantage using data from the Housing First controlled trial in Canada. Method: The study used a mixed method design. Latent growth curve and growth mixture models assessed the impact of working alliance across the sample as a whole and within subgroups with different patterns of housing stability. Thematic analysis explored the factors that may affect the quality of working alliances within different subgroups. Results: Three distinct trajectories of housing stability emerged (i.e., Class 1: “sharp, sustained, gradual decline”; Class 2: “hardly any time housed”; Class 3: “high, sustained, and gradual decline”) with professional helping relationships having different effects in each. The analysis revealed structural and individual circumstances that may explain differences among the classes. Conclusions: The findings underscore the role of the relationships, as distinct from services, in major interventions for highly disadvantaged populations, and draws new attention to the temporal patterns of responses to both the quality of relationship and targeted interventions.


Background
There is good evidence that the provision of housing can alter the trajectories of people facing severe and multiple disadvantage, particularly those experiencing homelessness.
Interventions like Housing First have been shown repeatedly to reduce homelessness, increase housing stability, decrease hospitalisation, and improve quality of life (1-6).
However, the effect sizes remain modest, and there is considerable variation in the reported impact on the different populations receiving such interventions (7).
One potential influence on that variation is the quality of the relationships between the workers who facilitate the transition into housing and the people facing multiple risks.
This relationship is typically part of an integrated support approach or case management -providing basic functions such as outreach, assessment, planning, linkage, monitoring, and advocacy -developed in the last three decades (see 8 for a review). Morse's review (9) highlighted several service and client factors that are critical components to effective case management. These include frequency of worker contact, lower lengths of time homeless, fewer psychotic symptoms, gender, and fewer substance abuse problems. Morse recommended that effective staff members need to have skills and abilities that enable them to develop trusting relationships with the people they support.
Relationships vary across many domains including those that comprise Bordin's concept of working alliance (10): agreement about tasks and goals and the nature of the bond between client and worker. Despite the multiple needs clients face, research has shown that people facing the greatest challenges seek first and foremost a personal connection with their workers (11)(12), facilitated by worker behaviours such as persistence, 'going the extra mile', being 'like a friend', and engaging in routine activities such as furnishing a home or going for a coffee (13)(14)(15). Practical help is also valued in these relationships as a way to further strengthen worker-client connections (12,16).
There is other evidence that the worker-client relationship can influence outcomes.
Goering, Wasylenki, Lindsay, Lemire, and Rhodes (17), for instance, found that a strong working alliance was a key element in achieving housing stability for 55 homeless and severely mentally ill clients connected to a hostel outreach program. Chinman, Rosenheck, and Lam (18) confirmed this in a sample of 2,798 homeless people with severe mental illness assigned to a program that offered outreach and intensive case management. Both research groups found that having an alliance was more beneficial than having none and that the strength of the alliance was inversely related to days of homelessness. However, Tsai, Lapidos, Rosenheck, and Harpaz-Rotem (19) found the strength and quality of the therapeutic alliance developed within the first three months of treatment in a supported housing program was not associated with any major housing and employment outcomes, although it did influence subjective reports of the quality of life and social support as reported by others (20).
Client and system characteristics play a role in the development of working alliances. In a qualitative study, Padgett, Henwood, Abrams, and Davis (21) found that individual circumstances -severity of mental illness -and system factors -pleasant treatment surroundings, worker acts of kindness, and access to independent housing-facilitated engagement with and retention in mental and substance abuse treatment services by formerly homeless psychiatric individuals. Conversely, co-morbid substance abuse, rules and restrictions, and the absence of one-to-one therapeutic support hindered service use. Klinkenberg, Calsyn, and Morse (22) found that being African American and not hostile while having greater perceived needs and more program contacts predicted strong alliances at month two in a sample of 105 individuals participating in an assertive community treatment for those with severe mental illness and high risk of homelessness.
Chinman, Rosenheck, and Lam (23) reported that having more social support, more subjective psychological problems, fewer overt psychotic symptoms, and more substance abuse problems were positively associated with building working alliances.
Several methodological limitations need to be noted, however (24). For instance, few of the studies described above included a control group. Follow-ups were typically short.
Most researchers treated their samples as homogenous whereas practitioners reported heterogeneity (25). Recent evidence using new analytical approaches such as growth mixture modelling (GMM) underlines the high variability and the complex patterns of change within this population (26).
The present paper has three aims. First, it explores the different patterns of change in housing stability over a two-year period in a population of homeless people randomly assigned to the Housing First programme in Canada. Second, it examines the impact of client-worker relationship as assessed via working alliances across the sample as a whole and within subgroups with different patterns of housing stability. Third, it explores the factors that may affect the quality of working alliances within different subgroups.

Methods
A mixed method design was used. The study's first two aims were assessed via quantitative analyses whereas qualitative analysis was used to explore the third aim.

PART I: Quantitative
Three research questions were explored: What is the impact of the working alliance on trajectories of housing stability among homeless people, controlling for age, gender, intervention, and adverse childhood experiences? In line with other research (17)(18)(19), it was hypothesised that the quality of the working alliances would have a positive influence on housing outcomes.
Are there multiple (unique subgroups of) developmental trajectories in housing stability among homeless people? As in previous studies (26)(27), a variety of housing stability trajectories were anticipated.
How does the working alliance affect existing unique subgroups of people experiencing homelessness?
In light of research underscoring the role of client and systemic factors in the development of relationships, for example substance abuse or access to independent housing, it was hypothesised that the working alliance might affect different subgroups differently.
Program description. The At Home/ Chez Soi randomised controlled trial of the 'Housing First' programme took place across five Canadian cities (identifier ISRCTN42520374). The study was approved by 11 research ethics boards. Participants, recruited between October 2009 and June 2011, were stratified by the degree of mental health service support required. Those with high needs were assigned to receive either assertive community treatment (ACT) or treatment as usual (TAU); those with moderate-needs to receive intensive case management (ICM) or TAU. Goering et al. (28) have described the study protocol in detail.
Participants. Participants were legal adults, absolutely or precariously housed, and had a mental disorder (with or without a concurrent substance use disorder). More than a third (34%) met the criteria for a psychotic disorder and two thirds (67%) for substance abuse or dependence. On a scale of 0-10, the average number of adverse childhood experiences was 4.6. Two thirds of the sample were male (67%), over half (57%) were aged between 35 and 54 years. Four-fifths were born in Canada and one fifth (22%) were aboriginal.
Nearly all (93%) did not work at baseline and fewer than half (45%) had completed high school.
Study procedure. ACT and ICM participants received housing of their choice as well as mental health services. TAU participants had access to local services. All intervention participants were supported by a worker through an ACT or ICM team; TAU participants were directed to other local resources, but availability was subject to local resources constraints. Comprehensive in-person interviews were conducted at baseline and every 6 months, and housing was assessed every 3 months. Qualitative interviews at baseline and 18 months were conducted with a tenth of the sample. Participants were compensated for their time with C$30 to C$50.
Measures. The Residential Time-Line Follow-Back Inventory (RTLFB) (see 29 for reliability and validity information) was used to measure the housing stability over time by the proportion of the number of days for which any type of residence (as living in one's own room, apartment, or house or with family) was available over the preceding 6 months.
Relationships between participants and workers were assessed with the Working Alliance Inventory Short Form, adapted (30) from the original 36-item version (31) which captured the degree to which clients and workers agreed on therapy goals and tasks and the quality of the bond between them. Each of the three sub-scales has 4 items rated on a 7-point scale (1=never, 7=always). Because the participants' accounts of their working alliances were assessed at multiple time points whereas the workers' perceptions were assessed once, only the client assessments were included in the analysis. Childhood trauma was assessed using the Adverse Childhood Experiences (ACE) Scale (32) which asks 10 questions related to childhood abuse and household dysfunction before age 18 years.
Quantitative analysis strategy. Data analyses were conducted using Mplus Version 7.4 (33) and missing data were handled using full information maximum likelihood (FIML) estimation. Prior to fitting growth models, preliminary analyses including individual trajectory plots, descriptive statistics and correlations were conducted using SPSS. Latent growth curve model. A well-fitting latent growth curve model (LGCM) was used to determine the overall sample trajectory.
LGCM assumes a single homogenous population in which individual variations around the overall mean growth trajectory are captured by the random intercept and slope coefficients (34)(35). Linear and quadratic growth models were used to find the best-fitting representation of change for the sample. Growth mixture model. The Latent growth mixture model (LGMM) (38) framework was used for data analyses to assess heterogeneity in patterns of change. The GMM approach adopted allows for differences in growth parameters across unobserved subgroups or classes (35) whereas LGCM assumes all individuals belong to a single population.
Additionally, LGCM assumes that covariates affecting class membership influence everyone in the same way. We hypothesised that qualitatively different subgroup trajectories may exist in the sample of homeless people reflecting variations in their health, impairment, and resilience. Progressively larger numbers of latent class (1-class to 4-class) solutions were run to determine the optimal number of classes (38)(39)(40). To determine the optimal class solution, a variety of fit statistics with classification accuracy were examined, so that average probability of belonging to the most likely class should be high, and the average probability of belonging to the other class should be low (41). In command was used to including covariates (age, gender, intervention, adverse childhood experiences) in the model to predict class membership. R3STEP results in less biased parameter estimates while maintaining a stable class solution and interpretable coefficients for the covariates (42). Working alliance was included in the model as a time varying covariate to explain variation in housing stability.

PART II: Qualitative
Qualitative Analysis Strategy. The quantitative analysis identified three patterns of associations between working alliance and housing stability. In groups one and three there was a strong relationship between working alliance and housing stability, and in group two only a weak relationship. The qualitative analysis focused on 10 participants from group one and 10 from group two.
Three sub-questions were examined: Do participants differ in their perceptions of relationship quality? Given the quantitative results, it was hypothesised that there would be group differences in definitions of good relationships.
Do structural problems in programme implementation explain variation in the results? It was hypothesised that structural factors might undermine the quality of relationships between workers and participants.
Do participant circumstances explain group differences in the results? It was hypothesised that the type and volume of adverse life circumstances and risks in group two, as compared to one, would limit the impact of worker-participant relationships on participant outcomes.
Participants from both groups were randomly selected from the available qualitative data pool. They varied with respect to age, intervention mode, gender, and number of adverse childhood experiences.

The analysis was informed by a critical realist approach (43) and employed Braun and
Clark's (44) six-stage thematic analysis approach. The analysis focused on the sentences in which the relationships with workers were mentioned. Codes were generated at each step. The analysis was conducted using the qualitative software NVivo 12. The trustworthiness of the emerging findings was tested twice. First, the author kept a reflexive journal in which noteworthy aspects of the analysis were recorded. Second, an external auditor (MSc student) scrutinised a third of the data and the emerging findings, generating further insights. For example, additional dimensions (e.g., care, honesty) valued by participants in good relationships were revealed by the audit.  were highly significant, indicating significant inter-individual differences in the rate of change in growth as a function of the covariates. The R 2 values (i.e., explained variance) in housing stability (baseline = 7.5%; six months = 35.2%; twelve months = 66.4%; eighteen months = 50.1%; twenty-four months = 34.2%) indicate that variation in housing stability is well explained by the growth factors and working alliance. The explained variance at baseline accounted for by only the intercept growth factor without the working alliance was small. Controlling for the effect of working alliance, the time-invariant covariates explained (R 2 = 14.2%, p = .660) in the intercept, (R 2 = 24.7%, p < .001) in the linear and (R 2 = 22.8%, p < .001) in the quadratic growth factors. Participants who received the intervention experienced a rise but also a decelerating growth in housing stability. The working alliance had a significant positive effect on housing stability at six months but the impact gradually declined from the twelfth month, although the declines were not significant.

Quantitative analyses
Growth mixture model of housing stability. Because the unconditional quadratic growth trajectory demonstrated acceptable fit to the data, it was decided to estimate the growth mixture modelling (GMM) to determine whether subgroups of individuals could be identified within the data. The latent class growth analysis (LCGA) used freely estimated class means but fixed within-class variances to zero, assuming within-class homogeneity.
Because estimating the GMM produced negative variances, growth factors variances were constrained to zero (46). Further, because intercept variance in the unconditional quadratic growth model was not significant and invariance of intercept means did not vary significantly across classes (χ 2 = 2.487, df = 2, p = .288), the intercept was constrained to be equal for all classes. Non-invariance in linear (χ 2 = 4738.001, df = 2, p < .001) and quadratic (χ 2 = 2249.947, df = 2, p < .001) growth factors showed that initial housing stability did not significantly differ between classes but that the rate of change and the decelerating change in growth trajectory were significantly different between classes.    Table 3 displays the growth parameters. Class 1 showed "sharp, sustained increase, gradual decline," with sharp initial increases and decelerating rates of change. Class 2 showed "hardly any time housed" and Class 3 "high, sustained, and gradual decline" in housing stability.

Predictors of class membership and within-class housing stability.
Once the growth mixture model was established, the R3STEP approach in Mplus was used to examine predictors of class membership. To include working alliance as a time varying covariate, housing stability was regressed on working alliance from T1 to T4 in each class.
The reference category was switched across regressions so that all pairwise comparisons were made. Age and intervention at baseline were significantly related to class membership with respect to the high, sustained, gradual decline and sharp, sustained, gradual decline classes (see Table 4).
In order to compare the means for age and intervention status across the classes, the DU3STEP approach, which assumes unequal means and variances, was used in a separate analysis. The results showed that older participants and those receiving the intervention were more likely to be in sharp, sustained, gradual decline, than younger persons or those not receiving intervention would be.* All other pairwise comparisons were nonsignificant.
The working alliance differentially predicted housing stability across the three classes. In the high, sustained, gradual decline class, the working alliance predicted housing stability at six (b = 0.304, p < .05), twelve (b = 0.278, p < .05), and twenty-four (b = 1.007, p < .001), but not eighteen (b = -0.011, p = .734) months. In the hardly anytime housed class, the working alliance did not predict housing stability over time, indicating that working alliance did not have an effect on housing stability for participants in this class.
In the sharp, sustained, gradual decline class, the working alliance significantly predicted housing stability at six (b = 0.228, p < .001) and twelve (b = -0.081, p < .05) months but not at eighteen (b = 0.001, p = .925) and twenty-four (b = -0.031, p = .674) months. The class varying results imply that class membership moderated the causal relationship between working alliance and housing stability when controlling for other covariates.

Qualitative analyses
The qualitative analysis closely examined 10 participant narratives from the sharp, sustained, gradual decline class (i.e., working alliance predicted housing stability) and 10 from the hardly anytime housed class (i.e., working alliance did not predict housing stability).
As previously described, three hypotheses were advanced to explain the differences in housing stability. The first was disproven by the qualitative analysis. Participants in both classes held similar perceptions of what counted as a good relationship with their worker.
They wanted those assigned to help them to be available when needed, able to listen, authentic, and caring as well as to provide practical help. They sought personal connections that went beyond professional transactions. The second hypothesis was supported by the qualitative evidence. Four structural factors appeared to explain the variation in relationship quality across the two groups: staff turnover, timing of worker assignment to participant, relational capability of the workers, and the uncertainty associated with the end of the programme (which may also explain variations in the growth curve analysis).
Frequent staff changes meant that some participants had to continually re-tell their personal stories. In some cases, this led to disengagement. I just need somebody and she tells me she's not qualified and that's it. I usually just don't answer the door, I peek out and see who it is and I don't answer the door.
Because why would I want to repeat the same story to a different person. (Female, 30, Intervention, 1 ACE, from sharp, sustained, gradual decline class) Comparing the quantitative records of worker assignments to participants with the latter's accounts of their relationships with those workers suggests that delay made it harder to form positive relationships.
Participants indicated that workers varied across sub-groups in their relational capability, including their ability to listen, provide practical help, and invest the time needed to form a connection that went beyond the minimum required by the service.
The end of programme was a continual threat to intervention participants in particular. It was a reminder that the relationship with the worker was to some extent a function of the intervention and could not endure, evoking feelings of sadness. This was compounded by uncertainty about their housing situation and support services. *In order to compare the means for age and intervention status across the classes, the DU3STEP approach, which assumes unequal means and variances, was used in a separate analysis.

Discussion
This study investigated the impact of professional helping relationships on the developmental trajectories over 24 months of housing stability for people facing severe and multiple disadvantage using a mixed-method design. Overall, the quality of the clientworker relationship made a significant contribution to the participants' housing stability.
The population was not homogenous. Three distinct trajectories of housing stability emerged (i.e., Class 1: "sharp, sustained, gradual decline"; Class 2: "hardly any time housed"; Class 3: "high, sustained, and gradual decline") with professional helping relationships having different effects in each. The analysis revealed structural and individual circumstances that may explain differences among the classes.
Not everyone had an equal chance of remaining stably housed: specific subgroups of people facing severe and multiple disadvantage face unique and different adversities so that interventions are not equally effective for different subgroups of homeless people. At the start, being male, the number of adverse childhood experiences, and being in the treatment-as-usual group reduced tenancy security. Older participants who received the intervention were more likely to remain in their homes over the study period. However, when the strength of the working alliance between the worker and participant was also considered, the above variables no longer predicted housing stability, demonstrating the importance of relationship quality.
There is overlooked heterogeneity in the recovery profiles of highly disadvantaged individuals (see 26 for exceptions). The first of three classes identified in this study had 1240 members (57.5% of the sample, with 73.6% of them receiving the intervention) (See Table 5). Participants in this group (who were older, disproportionately female, and with adolescent-onset homelessness) started with an initial spike in housing stability that was maintained until the last six months of the study when it diminished. The second group, representing over a third of the sample (37.1%, with 33.4% of them receiving the intervention), were the least successful in achieving housing stability and spent almost all the time unhoused. Participants in this group were predominantly male (73%), had longer histories of being homeless, could count on fewer people for social support, and spent a higher proportion of the study period imprisoned than did participants in the other groups.
The third group and smallest group (5.5%, with 53.3% of them receiving the intervention) followed a similar pattern of housing stability that was similar, but significantly different from, that of the first group. Participants were older when they first became homeless and spent fewer years being homeless, relied on larger social networks, and had higher rates of substance use than participants in the first group (comparable to that of participants in second group).
The non-linear patterns of housing stability may be important, as suggested by Adair et al. (26) too. Policy makers and intervention scientists think programmatically in one-or twoyear blocks, but the disadvantage that qualifies participants for intervention is accumulated over a lifetime. More attention to programme endings may be called for since it was a great source of anxiety for programme participants and contributed to housing instability for the most successful individuals in this study.
There was significant variation both in the impact of the intervention and in the contribution of the working alliance between support workers and participants.
Relationships in the first year predicted housing stability over time for one group (Class 3), in the first year for another group (Class 1), and not at all for a third group (Class 2).
The qualitative analyses revealed a number of possible reasons for this variation. In line with previous work (21,23), individual circumstances played a role in relationship development and housing instability. Participants' lifestyles (drug use) and systemic responses to that lifestyle (imprisonment) may make it harder for them to form relationships and easier for those relationships to be disrupted. For example, participants in Class 2 spent more time in prison than those in Class 1; this might have affected relationships by reducing the amount of time available for workers and participants to meet while also adding to housing instability.
Counter-intuitive impacts were found. A stable home can produce time for reflection, guilty feelings about the impact of past behaviours on family and friends, and a consequent backing away from these relationships and supports. Depletion and active avoidance of social connection can make it harder to form and sustain the effects of positive professional relationships. Padgett, Henwood, Abrams, and Drake (47) also found that people who experienced serious mental illness, substance abuse, and homelessness used a "loner talk" when talking about themselves in relation to others. However, Padgett et al. (47) also revealed that the participants had a desire to connect with others but were impeded by other factors, such as ambivalent nature of family ties, the focus on achieving housing stability rather than relationships, negative social networks, and substance abuse.
These findings underscored the role of the relationships, as distinct from services, in major interventions for highly disadvantaged populations, and draws new attention to the temporal patterns of responses to both the quality of relationship and targeted interventions.
This study also explained significant variations among sub-groups of participants and elucidated the characteristics of high-quality relationships between workers and participants, underlining the importance of a sense of personal connection (11-12, 14-15).
Study participants understood what counted as a quality relationship, but there was considerable variation in their (and their workers') ability to secure such relationships.
Temporal aspects of these relationships were also highlighted. The intervention -the provision of housing -endured but the change of workers was sometimes experienced by people as a sense of personal loss (47), and which contributed to housing instability.
There are several limitations to the study. The concept of the working alliance, although widely used, fails to fully capture the essence of worker-client relationships. Subcomponents of the working alliance were not identified, although this could have resulted in further understanding of how working alliances function in this population. Inclusion of different outcomes, in addition to housing stability, could have yielded a fuller picture of how relationships affect outcomes. Estimation of the mixture model did not use predefined group membership that could identify different trajectories for intervention and treatmentas-usual groups, although in the only study that has done this (26), the best performing class was similar to the one found in this study. The R3STEP analyses avoided biased class formation when including of covariates by retaining original class membership but the predictors' coefficients may have been biased the R3STEP procedure does not allow missing cases on exogenous predictors. Causal relationships between covariates and class formations should thus be inferred with caution. The themes provided by the qualitative analysis could have been further explored by increasing the number of transcripts analysed.
The study also had several strengths. It added to a body of evidence showing that professional helping relationships can have positive effects in severely disadvantaged groups (17)(18). The analysis overcame the methodological limitations associated with previous studies (e.g., 19) including the absence of control groups, brief follow ups, and assumptions of sample homogeneity. Sample selection for the qualitative analysis was driven by the quantitative analysis.

Consent for publication
Not applicable.

Availability of data and materials
The data that support the findings of this study are available from St. Michael's Hospital, Unity Health Toronto but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available. Data are however available from the authors upon reasonable request and with permission of St.

Figure 2
The class profile plot and Table 3 displays the growth parameters.