House Level Latent Classes as Predictors of Recovery and Evictions

Abstract

The current study explored whether substance abuse recovery houses could be categorized into meaningful classes which might be associated with house evictions as well as changes in individual-level recovery capital. A total of 602 individuals from 42 recovery homes were followed for up to 6 data collection periods over 2 years. House level latent class analyses were based on house level data. A 3-class model fit very well (entropy 0.94) and better than a 2-class model. Class profiles examined concurrent (averaged across waves 1 and 2) house and resident level variables (e.g., gender, race, age, employment, education). Class was then used to prospectively predict outcomes of the hazard of eviction and improvement in a recovery index over waves 3–6. One latent class representing 45% of the recovery houses had the highest density of members willing to loan, able to pay their rent, active involvement in outside chapter activities--this group of houses had the best outcomes including the lowest eviction rate and highest mean recovery factor. The two other classes had higher eviction rates, with one having the lowest density of friendship, selectivity of residents, and ability to pay rent. The other of the higher eviction-rate classes surprisingly had the highest density of friendship and advice seeking, but the lowest density of willingness to loan. These findings suggest that there are meaningful differences in types of recovery homes, and that house characteristics appear to influence recovery changes and eviction outcomes.

It is estimated that each year over 250,000 individuals with substance use disorders reside in recovery homes, and many of residents are at risk for a variety of health and mental health disorders, and they often have high prior rates of criminal justice involvement (Jason, Wiedbusch, et al., 2020). There are two general types of recovery homes, those that have staff and those that are self-run. In the latter, with residents manage membership issues, rule enforcement, behavioral management, and house finances. Outcomes have generally been positive for both types of sober living settings. For example, at a six month follow up, those in more traditional staff run sober-living homes in California reported 51% abstinence from substances (Polcin & Henderson, 2008).

The Oxford House (OH) network (Oxford House, 2020) is the largest organization of resident-run recovery homes in the US. Studies of these homes have also shown a beneficial effect for residents (Author) relative to comparable individuals who did not enter a sober living environment following acute treatment for addiction. However, a recent study by our group based on a relatively large sample of 42 OHs from three different geographical regions of the US found that even within the OH network, a significant amount of variation in rates of house eviction (for failure to remain sober, rules violations, etc.) appears to depend on house-level factors, rather than individual characteristics (Jason, Guerrero et al., 2020). This finding raises questions as to how OHs may differ, i.e. on what characteristics, and whether there are parsimoniously identifiable classes of OHs. If such classes could be identified, their characteristics might suggest why some OHs are more successful than others. A class structure could also be used to explore environment-contingent mechanisms of recovery.

We can think of these environmental effects as potential differences in social capital, as it relates to supporting residents’ recovery. Over the past 15 years, a variety of prominent researchers including Best and Laudet (2010) and Cloud and Granfield (2008) as well as federal agencies (SAMHSA, 2012) have linked Social Capital Theory to the recovery process. Empirical work with social networks have demonstrated the promise of utilizing this theory to those with substance use disorders (Kelly, Stout, Greene, & Slaymaker, 2014; Patterson et al., 2020; Tracy et al., 2012).

Recovery homes are intended to help residents transition from new abstinence to longer term recovery. Recovery capital, conceptualized as a recovery-supportive form of social capital specific, is a useful way to link Social Capital Theory with empirical indicators of social relationships and the resources they may provide access to (Best & Laudet, 2010). While living in a recovery home, residents can both make use of recovery resources and generate them (via social or instrumental support) for other members. The social networks that emerge in these settings involve multiple types of relationships, which may be affected by both resident and house characteristics.

Little is known about what house contextual or interpersonal factors explain relationship formation in recovery homes. In addition, knowledge is still limited regarding the process of social transition from active substance use to recovery. While the effects of affiliation with a recovery group (e.g., becoming an recovery home resident) are well described by extant studies, the process of such affiliation is much less well-understood. Indeed, although Humphreys et al. (1994) noted several decades ago that “…the process by which persons become affiliated with mutual help groups is a topic of interest” (italics added), very little empirical research has addressed this fundamental question since. But interest in this transitional process continues. For instance, the “Social Identity Model Of Recovery” (SIMOR; Best et al., 2016) has identified identity transition as an essential ingredient of successful recovery, in which the “addict identity” operative during active substance dependence is replaced by a new “recovery identity”. The theoretical perspective explaining identity transition is expressed in terms of a change in reference group(s) (e.g McCall & Simmons, 1966). This theoretical perspective maintains that humans construct an identity (“who I am”) from the reflected image they receive from others in their social environment.

In a recent study, Author reported that “friendship” and “willingness to lend” are more common and reciprocated than “advice seeking”. They visually represented as graphs for each network type—friendship, willingness to loan, and advice-seeking—in order to identify broad structural patterns. The houses were then classified in terms of their density on each of the network types, as high, medium, or low. The resulting nine categories (low, medium, and high density for each of friendship, willingness to lend, and advice-seeking networks) found variation in social network structure. However, other house-level characteristics were not considered, and classification was descriptive only, and did not utilize any statistically based classification methodology.Like most recovery programs, recovery homes do not work for everyone. For example, one study found that only about half of new residents in Oxford House (OH) recovery homes remained in residence long enough to achieve a sustainable level of recovery (Jason & Ferrari, 2010). The recovery influence transmitted by the house culture may depend on the quality of a resident’s house support network as well as other individual-level factors such as recovery status. The purpose of our study is thus to understand which types of houses most facilitate recovery, and under what circumstances. The current study attempted to categorize the same set of recovery houses using Latent Class Analyses (LCA), and using a broad set of class of recovery capital-related criteria as class indicators for waves 1–2 of the study. The classes thus derived were described in terms of their differences on class indicators, and further explored by examining class differences in additional concurrent house and resident level variables (e.g., gender, age, race, employment, education, time in residence prior to joining the study, etc.). Finally, the classes were employed to prospectively predict across waves 3–6 the distal outcomes of the hazard of eviction and growth of an index of individual recovery. These exploratory analyses were intended to further our understanding of complex individual–environment domains among those with substance use disorders residing in recovery homes.

Method

Settings

The study was conducted in a set of self-run Oxford Houses (OH), which are rented and gender-segregated, housing about 6 to 12 individuals in recovery. All residents are required to follow three primary rules; paying their fair share of the rent (which is usually from $100 to $125 per week), contributing to the maintenance of the home, and abstaining from using alcohol and other drugs.

Data were collected from OHs located in North Carolina, Texas, and Oregon. Member-elected house presidents were asked to introduce the study to residents by reading a project-provided script about the study; houses were accepted into the study if the house president and all or all but one member agreed to participate. The first thirteen consenting houses from each state were accepted, and three more houses were added for a total of 42. One house dropped out of the study, replaced by another house after wave 1 brings the total to 43 houses. However, only 42 houses had 2 or more waves of information available on their residents. Once a house was recruited, all residents of those houses were invited to participate in the study by one of three recruiters who visited houses to explain the study to possible participants.This longitudinal study was designed to collect information from participating house members every four months over a 2-year period for a total of 7 waves. This research design involved a) new residents entering the recovery homes and participating in the study after the initial wave 1 assessment, b) residents who declined to participate at wave 1 but joined the study at a later wave and c) residents that left over the course of the study. Details on how this design affected statistical models are presented below (see the analytic approach section).

Participants

Participants were recruited and interviewed by field research staff in face-to-face meetings. Participants completed measures of their demographics, stress, self-esteem, support and social networks. They were compensated $20 for completing their assessment. All participants provided written consent after being given an explanation of the study by one of the three recruiters. Permissions for this study were granted by the DePaul University Institutional Review Board.

There were 714 residents of the OHs during this period of time, of which 666 (93%) agreed to participate in this investigation. Of the 666 who agreed to participate, 64 were excluded for the following reasons: 34 entered an OH for the first time at wave 7 and we were not able to determine whether they relapsed given that the last data collection occurred during wave 7; 15 were forced out because their OH closed before wave 7 and before they filled out a second survey; 15 were missing their reason for leaving and had only filled out 1 survey before leaving. The final study/analysis sample was 602 - (84%) of the population.

The analysis sample of 602 was 51% male, with a mean age of 37.0 years (SD = 10.5). Participants identified as White (78.8%), Black (8.5%), Latinx (10.0%), with all other ethnicities accounting for 2.7% (Asian American, Alaskan Native, American Indian, and Pacific Islander). Missing data on exits and predictors were minimal, never exceeding 5% of the 602 residents for any variable.

Resident demographic information included age, sex, race/ethnicity, education, employment and time in residence prior to joining the study. Race/ethnicity was broken down in to 4 categories, White (78.8%), Black (8.6%), Latinx (10.1%) and all other (2.5%). Oxford Houses are gender-specific and accordingly, gender was included as a house level predictor. We evaluated residents’ employment status in 3 ordered categories of 1) unemployed or other forms of income (disability, student, retirement), 2) part time and 3) full time. We classified resident educational attainment in 5 ordered categories as 1) less than high school, 2) high school graduate or GED, 3) technical degree, 4) some college, and 5) college graduate or higher.

Measures

Recovery Factor

The recovery factor was based on a confirmatory factor analysis of 8 recovery capital indicators (Jason et al., 2021). Confirmatory factor analysis had been used to examine eight measured indicators (self-esteem, hope, quality of life, stress, social support, self-efficacy, sense of community, and wages) and a 1-factor model of recovery emerged with a good fit (RMSEA = .04). This factor allows an assessment of the progress an individual is making in treating their substance use disorder. In other words, this recovery factor combined several variables that represent individual and group-level elements of recovery, to capture the complex individual and environmental resources which contribute to the recovery capital of those with substance use disorders. Recovery-supportive domains like these may be key to understanding important recovery dynamics, and are comparable to a recovery capital construct that taps resources within SAMHSA (2012)’s 10 guiding principles of recovery. This analysis included all recovery factor scores available for each resident during their first residency period up to wave 6 of the study (comparable to that reported in Jason et al. 2020). The estimated recovery factor scores were calculated from the following measures as indicators:

Wages.

The square root of the wages for the last 30 days was computed and used as a continuous variable.

Quality of Life.

The World Health Organization Quality of Life Assessment-Brief (WHOQOL Group, 1998) is a 24-item questionnaire that assesses quality of life across four dimensions: social relationships, environment, physical, and psychosocial. This scale has been validated in substance using populations (Garcia-Rea & LePage, 2010). The subscales varied in their reliability (αs = .89 for social relationships, .84 for environment, .83 for physical, and .83 for psychological). The alpha for the measure for our sample was .89.

Self-efficacy.

The Drug Taking Confidence Questionnaire (Sklar et al., 1999) is an 8-item survey measuring self-efficacy in terms of abstinence. Participants are asked to consider themselves in 8 theoretical high-risk situations and indicate how confident they are that they could resist the urge to use a substance given the theoretical circumstances. This measure for our sample has good reliability (α =.95).

Self-esteem.

The Rosenberg’s 10-item, Self-Esteem Scale (Rosenberg, 1965) was utilized to measure the participant’s positive and negative feelings about the self. SES is a widely used 10-item, global self-esteem scale measured on a 4-point Likert Scale ranging from “strongly agree” to “strongly disagree”. Examples of items include “I think I have a number of good qualities,” “I take a positive attitude towards myself,” and “I feel I do not have much to be proud of”. The internal reliability of the SES scale for our sample was .92.

Stress.

The 4-item Perceived Stress Scale (Cohen et al., 1983) measured the degree to which situations in participants lives are appraised as stressful. Examples of items include “how often have you felt that you were unable to control the important things in your life?” and “how often have you felt difficulties were piling up so high that you could not overcome them?” The internal reliability of the perceived stress scale for our sample was .73.

Social support.

The 12-item Interpersonal Support Evaluation List (Cohen & Wills, 1985; Cohen, Mermelstein et al., 1985) was utilized to measure three types of perceived social support (tangible, appraisal, and belonging). Tangible support refers to instrumental aid and monetary assistance; appraisal support refers to having someone to talk to about one’s problems; and belonging support refers to the availability of people one can do activities with. The Interpersonal Support Evaluation List consists of 12-items measured on a 4-point Likert scale ranging from definitely false to definitely true. The internal reliability of the support scale for our sample was .88.

Sense of Community.

The Psychological Sense of Community (Jason et al., 2015) is a 9-item scale utilized to measure participant’s sense of community at their Oxford House. Examples of items include “This Oxford House is important to me” and “For me, this Oxford House is a good fit”. The three subscales are Entity, Membership, and Self, and for our sample, they have Cronbach alphas of .67, .92, and .91, respectively. The Sense of Community (SOC) scale was used as a whole measure (α = .91).

Hope.

The State Hope Scale (Snyder et al., 1996) consists of 6 items that measure participants’ current state of hope. The Hope measure contains two sub-scales Agency (α = .94) and Pathways (α = .81). We included a 3-item subscale of hope that measures Environmental Context (Stevens et al., 2014) (α = .97). This 9-item hope scale was analyzed as a whole measure, and for our sample the α = .90.

Home Process Variables

The house presidents at study wave 1 filled out the Recovery Home Processes Questionnaire, providing data on the following variables.

House new applicant acceptance rate was assessed by this question: “Out of every 10 applicants, how many does your house accept?” Higher scores indicated more acceptance. We named this variable Selective with the scale reversed so that higher scores indicate more restrictive policies toward acceptance, which we considered a more positive policy.

House average resident poverty was assessed by this question: “On average, how many residents have difficulty meeting their monthly financial obligations?” Higher ratings signified being more impoverished. We named this variable Affluence which was reverse scored so higher scores are considered better.

House involvement in chapter activities was assessed by this question: “How involved is the house in chapter activities: Not at all involved, somewhat involved, much involved, a great deal involved.” Higher scores on this variable named Active indicated more involvement.

House savings was based on this question: “How would you describe the financial condition of the house: significant debt (>$5,000); some debt ($1–5,000) neither debt or savings; some savings ($1–5,000); significant savings (>$5,000)”. This variable was named Savings, with higher scores indicating more house savings.

A final house level variable was the sex of the house (female or male).

Eviction Outcomes

We also included the Addiction Severity Index-Lite (ASI-Lite; McLellan et al., 1997) to assess evictions for problems in drug and alcohol use over the past 30 days. The ASI-Lite has been demonstrated to have good validity and reliability (Cacciola, et al., 2007). This instrument along with other information from the house officers was used to create the following exit status variables: positive outcomes (46% left on good terms), negative outcomes (31% either relapsed, or left for disruptive behavior or financial reasons), still a resident (22%), or left with no reason available (1%). In this study, the 0/1 indicator variable created from this information was left for an eviction.

Relationship types

The definition of a tie for each relationship has been operationalized and defined previously (Author). In brief, a friendship tie means the resident doing the rating (hereafter termed “ego”) considers a rated housemate (hereafter termed “alter”) as a friend or close friend, a loan tie means ego is willing to loan $100-$500 to alter and an advice seeking tie means that ego goes to alter very often or often for advice. The density within a house is the proportion of all ties that exist divided by all ties that could exist. The densities are proportions that range from 0–1 and indicate more or less ties of each type within a house. High friendship density means that most people within a house are friends or close friends. High loan density means that most people within a house would be willing to lend $100-$500 dollars to a housemate. High advice seeking density means that most people within a house were very often or quite often go to a housemate for advice.

Statistical Analysis

House level Latent Class Analyses (LCA) were utilized to identify distinct groups of recovery houses. The LCA were based on house level data at waves 1 and 2 and the LCA indicators included average network density across waves 1 and 2 of friendship, advice seeking and willingness to loan; House president ratings at wave 1 of 1) house new applicant acceptance rate (Selective), 2) typical house resident financial status (Affluence), 3) house involvement in chapter activities (Active), and 4) house savings (Savings).

We started with the standard assumptions that within classes, the indicators are multivariate normally distributed and the within class covariance matrix is invariant across the classes and diagonal (i.e., zero correlation, so called local independence assumption). The correlation between friendship and advice seeking densities, however, was high enough to warrant testing the local independence assumption. Preliminary models indicated that the residual, within class correlation was still strongly significant in both 2 and 3-class models so it was retained for all models. House savings did not discriminate the classes in either the 2 or 3-class model and thus was dropped as a class indicator. Due to the limited sample size of houses, we did not consider more than 3 latent classes.

To simplify interpretation of latent class profiles, the indicators were standardized prior to LCA to means of zero and standard deviations of 1. After model fitting, class mean profiles were further standardized using the estimated within class covariance matrix so they represent deviations from the overall population mean in residual, within class standard deviation units (similar to Cohen’s d). We explored both 2 and 3 class solutions, as preliminary Bayesian versions of both of these models are acceptable by goodness of fit criteria (p = .19 for 2-class and p = .62 for 3-class).

After class extraction, we obtained estimates of latent class membership for each house, so called pseudo-classes, and profiled the classes using additional concurrent (averaged across waves 1 and 2) house and resident level variables (gender, age, race, employment, education, prior time in residence, evictions, and the recovery factor). This was followed by using the wave 1 and 2 house level pseudo-classes to prospectively predict distal outcomes of the hazard of eviction and growth of the recovery factor over waves 3–6.

Typically, analyses involving predictors of latent class membership or effects of latent class membership on distal outcomes are either conducted in 1 step using all the predictors and outcomes together in a simultaneous analysis or in 3 separate steps. We used a simplified version of the 3-step approach. The first step in a 3-step analysis is to decide on the number of classes, the second step is to generate estimates of latent class membership (pseudo-classes), and the third step brings in the predictors or the distal outcomes of pseudo-class but using bias corrections to compensate for the misclassification error when assigning subjects to pseudo-classes. We used the 3-step procedure but omitted the corrections for misclassification error because it was negligible (see results below).

The hazard model of evictions was a 2-level (residents within houses and houses) discrete time hazard analysis that in addition to the pseudo-classes included a linear baseline hazard and control variables of prior time in residence and the recovery factor and initially included a random intercept for house level differences in evictions. The growth model for the recovery factor was a 3-level growth model (repeated measures within residents, residents within houses and houses) that in addition to the pseudo-classes included prior time in residence as a control variable and initially included random intercepts and slopes at both the resident and house level. Both models for distal outcomes were tested for cohort effects based on when a resident entered the study but none were detected. We simplified the distal outcome models by eliminating random effects that were not significant except we retained at a minimum a random intercept at either the resident or house level regardless of whether it was significant or not.

Results

All the indicators had at least 1 significant class contrast in the 2- or 3-class model. The 3-class model fit substantially better than the 2-class model (See Table 1) as indicated by information criteria [AIC, BIC or sample size adjusted BIC (aBIC)]. Entropy was very good for both the 2- and 3-class models, .90 and .94 respectively, on a scale of 0–1 with 1 being perfect class prediction.

Table 1.

Model comparison of two vs. three latent classes.

Classes	Parameters	AIC	aBIC	BIC	Entropy
2	20	684.10	656.22	718.85	0.90
3	27	664.17	626.54	711.09	0.94

2-Class Model

Figure 1 presents the mean profile plot for the 2-class model and there are prominent differences on each indicator. Those houses that are in Class 1 have higher densities of friendship and willingness to loan, as well as higher member selection criteria, more resident affluence, and higher involvement in chapter activities (see Table 2). The only indicator where Class 2 has a higher mean is the density of advice seeking. Several prior articles have suggested that high levels of advice seeking are related to less positive outcomes (Author).

Figure 1:

Two Class Solution

Table 2.

^abc Class differences among indicators for the 2 Models

2-Class Model
indicator	contrast	est	se	z	p	std est
friendship	class 1 - class 2	0.450	0.286	1.573	0.1156	0.463
advice	class 1 - class 2	−0.799	0.265	−3.015	0.0026	−0.872
loaning	class 1 - class 2	1.084	0.292	3.712	0.0002	1.296
selectivity	class 1 - class 2	0.351	0.343	1.023	0.3062	0.356
affluence	class 1 - class 2	0.942	0.284	3.317	0.0009	1.063
activities	class 1 - class 2	1.033	0.280	3.689	0.0002	1.202
3-Class Model
indicator	contrast	est	se	z	p	std est
friendship	class 1 - class 2	1.205	0.382	3.154	0.0016	1.721
advice	class 1 - class 2	0.015	0.287	0.052	0.9583	0.027
loaning	class 1 - class 2	0.822	0.282	2.915	0.0036	1.045
selectivity	class 1 - class 2	0.657	0.313	2.099	0.0358	0.697
affluence	class 1 - class 2	1.607	0.486	3.307	0.0009	2.240
activities	class 1 - class 2	0.871	0.657	1.326	0.1849	0.964
friendship	class 1 - class 3	−0.645	0.320	−2.016	0.0438	−0.921
advice	class 1 - class 3	−1.941	0.252	−7.702	0.0000	−3.597
loaning	class 1 - class 3	1.481	0.298	4.970	0.0000	1.883
selectivity	class 1 - class 3	−0.158	0.410	−0.385	0.7000	−0.168
affluence	class 1 - class 3	0.377	0.325	1.160	0.2460	0.526
activities	class 1 - class 3	0.853	0.382	2.233	0.0255	0.944
friendship	class 2 - class 3	−1.850	0.318	−5.818	0.0000	−0.921
advice	class 2 - class 3	−1.956	0.258	−7.581	0.0000	−3.597
loaning	class 2 - class 3	0.659	0.256	2.574	0.0100	1.883
selectivity	class 2 - class 3	−0.815	0.341	−2.390	0.0168	−0.168
affluence	class 2 - class 3	−1.229	0.527	−2.332	0.0197	0.526
activities	class 2 - class 3	−0.019	0.479	−0.040	0.9684	0.944

^aThe est values are the differences in the raw indicator means across the contrasted classes.

^bThe std est values are the differences in the standardized indicator means across the contrasted classes.

^cThe class profile plots are of standardized class means.

Profiling the 2-Class Model

The following house or house level averages from waves 1 and 2 of resident level variables were significantly different across the 2 classes: The recovery factor, evictions, and prior time in residence. Class 1 was higher on the recovery factor and had more prior time in residence. Class 2 was higher on evictions. Gender, age, race, employment, education and positive exits were not significantly different across the 2 classes.

Distal Outcomes in the 2-Class Model

Within the growth model for the recovery factor, class 2 had a significantly lower initial starting point at wave 3 compared to class 1 but class effects on the slope were not significant and thus, the relative differences between the classes at the wave 3 starting point were maintained over time until the end of the study at wave 6. Within the hazard model for evictions, across waves 3 to 6, class 2 had a significantly higher hazard of eviction compared to class 1.

3-Class Model

Figure 2 presents the indicator profile plot for the 3-class model. Although not apparent from the profile plot, cross-tabulating pseudo-classes from the 2- and 3-class models showed that Class 1 from Figure 1 (the 2-class model) remained almost identical in Figure 2 (the 3-class model) except for one house moving to another class. In contrast, Class 2 from the 2-Class model split in to two new classes (classes 2 and 3) in the 3-class model and Class 3 (the smallest class) bears little resemblance to either of the classes from the 2-class model. Class 2 from the 3-class model, however, still bears some resemblance to Class 2 in the 2-class model.

Figure 2.

Three Class Solution

In the 3-class model, the relation of Class 1 to Class 2 is about the same as in the 2-class model except for the mean density of advice seeking which was about the same in both classes (see Table 2). Class 3, however, had the highest mean densities of friendship and advice seeking compared to classes 1 and 2 but the lowest mean density of willingness to loan. Class 3 was similar to class 1 on selectivity and affluence but similar to class 2 on involvement in chapter activities.

Profiling the 3-Class Model

The following house or house level averages from study waves 1 and 2 of resident level variables were significantly different across at least 2 classes: The recovery factor, evictions, prior time in residence and employment status. Class 1 and class 3 were higher than class 2 on the recovery factor; class 2 and class 3 were higher than class 1 on evictions, class 2 and class 3 were lower than class 1 on both prior time in residence and employment status. No other class contrasts were significant and gender, age, race, education and positive exits were not significantly different across any of the classes.

Distal Outcomes in the 3-Class Model

Within the growth model for the recovery factor, class 2 had a significantly lower initial starting point at wave 3 compared to class 1 and class 3 but class effects on the slope were not significant and thus, the relative differences between the classes at the wave 3 starting point were maintained over time until the end of the study at wave 6. Within the hazard model for evictions, across waves 3 to 6, class 1 had significantly lower hazards of eviction than class 2 or class 3.

Discussion

Our study found that we could effectively classify a group of recovery homes using either a 2-class or 3-class model (see Figures 1 and 2). The 3-class model fit better than the 2-class model but both had an acceptable fit. The latent classes were based on house level characteristics derived from the first 2 waves of our longitudinal study including network densities of a) friendship, b) advice seeking, and c) willingness to loan and house president ratings of a) house financial savings, b) resident affluence, c) involvement in chapter activities and d) selectivity with respect to new, prospective residents. The class mean profile plots are shown in Figures 1 and 2 for the 2- and 3-class models respectively and show how the latent classes are defined by the observed class indicators. Class 1 (solid black line in Figures 1 and 2) was essentially same group of houses in both the 2- and 3-class models. Class 2 in the 2-class model split to derive classes 2 and 3 in the 3-class model. Class 1 in the 2-class model would appear to be the class that would tend to have better outcomes because with the single exception of the density of advice seeking, class 1 has higher class indicator means than class 2. Our prior work on the relation of advice seeking with resident outcomes, however, suggests that high levels of advice seeking may be associated with poorer resident outcomes. The relative relation of class 1 and 2 in the 2-class model is approximately maintained in the 3-class model but the relation of class 3 to the other 2 classes is more complex and difficult to describe. Class 3 has substantially higher mean levels of friendship and advice seeking than classes 1 and 2 but the lowest level of willingness to loan. Selectivity and affluence resemble class 1 but involvement in chapter activities resembles class 2.

In summary, the LCA suggests that regardless of whether we allow for 2 or 3 classes, about 45% of the houses belong to a class (class 1) that generally has the best outcomes, a lower rate of evictions and higher levels of the recovery factor. For the remaining 55% of the houses, the eviction rate is substantially higher regardless of whether we allow for 2 or 3 classes. Among these 55% of houses, however, the 2- vs 3-class models are differentiated by the level of the recovery factor. In the 3-class model, a smaller subclass, class 3 (25%) splits off from class 2 in the 2-class model and this class resembles the high functioning class 1 on the recovery factor.

These findings are complex but they certainly highlight the need to better understand how the individual in early recovery detaches from old, substance-using reference others and forms new attachments to recovery-supportive others. The SIMOR model is based on a conception of recovery that includes not only sobriety, but also changes in lifestyle and world-view that support sobriety in the recovering individual’s social settings. The focus on social identity explicitly directs attention to factors that support the individual’s progress in constructing this new identity. Best and colleagues (2016) go on to argue that social identity has long been recognized as being strongly linked to the lifestyles and world-views of one’s social relationships, that is, of the reference groups one is part of (e.g., Turner, 1978; Walker & Lynn, 2013).

Our study helps better understand these types of transitions in different types of recovery housing. After studying mean profiles on the class indicators, we tried to further elucidate the nature of the latent classes by looking at the concurrent relations between latent class membership and a range of house and house averages of resident level variables from waves 1 and 2 including gender, race, age, prior time in residence, house eviction rate, and a recovery capital construct that captures 8 important dimensions of the recovery process (Author). For the 2-class model, class 1 was higher on the recovery factor, had more prior time in residence and was lower on evictions. For the 3-class model, class 1 and class 3 were higher than class 2 on the recovery factor; class 2 and class 3 were higher than class 1 on evictions; class 2 and class 3 were lower than class 1 on both prior time in residence and employment status.

To prospectively validate the latent classes, we tested the predictive relation of latent class membership derived from waves 1 and 2 against distal outcomes in waves 3–6 that included the hazard rate of house evictions and the growth of the recovery factor. In the 2-class model, class 1 had a lower hazard of evictions and a higher initial mean at wave 3 in the growth model for the recovery factor. Slopes were not different so the relative difference between class 1 and 2 was maintained to wave 6. In the 3-class model, class 1 had a lower hazard of eviction during waves 3–6 than classes 2 and 3. Class 1 and 3 both had a higher initial starting point at wave 3 for growth of the recovery factor compared to class 2. Mean slopes were not different across the classes so the relative differences between class 1 and class 3 vs class 2 were maintained to wave 6. From a theoretical point of view, it is quite possible that those living in class 1 housing might have identified their transition to new reference others as the lynchpin of recovery (Best et al., 2016). Nevertheless, our study is only able to outline some possible predictors of this transition, due to a lack of relevant empirical studies.

Perhaps the most challenging class of houses to understand was Class 3 in the 3-class model. This class had elevated friendship and advice seeking density, whereas loaning was at the lowest level. Other house characteristics did not really differentiate Class 3 from the optimal Class 1. It is therefore puzzling why evictions were so high in Class 3 houses because they don’t look much different from Class 1 on the recovery factor or the house president ratings (in fact, Class 3 had higher mean house savings than Class 1). Perhaps being less willing to loan money (possibly representing low trust) despite claiming to be very friendly represents a risk factor. Among the reasons for eviction, 4% of those were evicted in Class 1 for not being able to pay their rent, whereas this figure increased to 16% in Class 3. Perhaps houses that are less willing to loan resources are also concerned about other aspects of their members such as ability to pay their rent. This might also include concerns that some residents are asking too many questions, as advice seeking was highest in this group of houses. Other findings also have pointed to risks involved in advice seeking. For example, Author found higher density of advice-seeking relationships was related to more stress, suggesting that residents go to others for guidance during stressful times. Author found higher density of “advice seeking” was predictive of greater relapse. High scores could be a marker for stress or a marker for social isolation and lack of trust. One of the benefits of LCA is that it easily allows for class membership to depend on such non-additive combinations of indicators.

The study’s findings complement those of Author who found the OH-level latent recovery factor as a strong predictor of average rates of relapse across homes, whereas the individual-level factor did not predict individual variation in relapse rates. In practical terms, this means, for example, the probability of relapse in any given time period is primarily related to the average of the “recoveries” of their OH colleagues, and not of their own personal “recovery” status. This is consistent with the idea that recovery can be strongly influenced by others with whom the recovering individual has regular social contact. That study as well as the current one emphasize the importance of house level variables, and certainly the current study suggests that houses can be differentiated into those facilitating member recovery capital and remaining in the houses.

We have a host of important and largely unexplored questions to address. Among these are there conditions where mentoring and friendship relationships can co-exist? What is the marginal value of additional mentors, friends, or close friends beyond one? What is the marginal risk of the number of substance users in one’s personal network? Does time in house have a dual effect, such that up to an (unknown) threshold time T, one is more likely to acquire a mentor, while after T, one is more likely to become a mentor? Is there a “substitution effect” such that having a fellow house resident as a mentor or having a mentor in one’s personal network outside the house confers the same benefit, or is a mentor more beneficial in one or the other location? These are but a few of the exploratory issues we intend to address in future data collection efforts. Recovery home social integration represents a tractable way to study this more general problem of how recovery supportive social relationships are acquired and maintained.

There are several limitations in this LCA study. It is unclear whether these different groupings of houses have differential longer term effects on residents once they departed from their recovery homes. The findings might be relevant only to those who are members of democratic, self-governing recovery homes like OHs, rather than other recovery homes that are staff led. In addition, the president ratings of house characteristics are relatively simple one item stems and their reliability and validity are still not known. Finally, the sample size was relatively small, and the number of houses within several of the classes in the 3-Class model were particularly small.

Our research might ultimately help patients in making better choices as well as social work clinicians in providing evidenced-based advice. For example, given the post-treatment psychosocial profile of a person in recovery (e.g., likelihood of obtaining work, etc.) who is bound for a recovery residence, our research might suggest ways to maximize the chances of successful recovery. In addition, our work helps define or describe a post-treatment trajectory with a greater likelihood of success and with a larger opportunity set. However, there are considerable implications of this research for more traditional substance use treatment facilities, for there is considerable variability in recovery, and if we can better understand the mechanisms, even those who are in these treatment settings could potentially develop better paths and outcomes for the clients.

Our study suggests that Social Capital Theory and the SIMOR model have much to help social workers and other community-based professionals understand the dynamics of how social relationship affect houses as well as individual recovery outcomes. Recovery capital is a complex process, but crucial for uncovering how recovery is promoted, and differential resident experiences in these settings, and how insights learned can be expanded to other types of settings. By identifying house types through which social environments affect health outcomes, our work can lead to better outcomes for those in recovery homes, and also potential help social workers restructure and improve other recovery settings. Recovery capital provides us with insight regarding house structure and dynamics, that might help addiction and social work professionals predict resident recovery trajectories. Such a framework can help the field of social work better understand network dynamics in recovery homes, and this could contribute to further theoretical development and investigation of recovery homes.