Context Matters: Home-level But Not Individual-Level Recovery Social Capital Predicts Residents’ Relapse

Abstract

The purpose of this study is to contribute to the literature on the prediction of substance use relapse, using sophisticated systems’ approaches to individuals and their contexts. In the current study of 42 recovery homes, we investigated the construct of social capital from the perspective of both recovery home residents and the house level. A confirmatory factor analysis found a latent recovery factor (including elements of recovery capital, comprising resources such as wages, self-efficacy, stress, self-esteem, quality of life, hope, sense of community, and social support) at both the individual and the recovery house level. Next, using longitudinal data from homes, an individual’s probability of relapse was found to be related to house rather than individual-level latent recovery scores. In other words, an individual’s probability of relapse was primarily related to the average of the “recoveries” of his or her recovery home peers, and not of his or her own personal “recovery” status. The finding that resident relapse is based primarily upon the total recovery capital available in the homes highlights the importance of the social environment for recovery.

Introduction

The field of community psychology has emphasized ecological approaches to measure complex individual-group transactions (Jason & Glenwick, 2016) where individual behaviors are understood as adapting to a social environment, and where this social environment involving a person’s social relationships also changes and adapts. In the field of substance use, this perspective would encourage the adoption of a multi-dimensional framework where social relationships are conceptualized as co-evolving over time, affecting and affected by recovery-related attitudes, behaviors, and personal networks (Campbell, Ash, & Bauer, 2008). For example, the Substance Abuse and Mental Health Services Administration (SAMHSA, 2012) updated its framework of recovery to a more holistic approach including both individual and more contextual components such as wellness, purpose, self-esteem, hope, self-efficacy, financial stability, social support, community, and having a stable and safe home. This framework further acknowledges that the process of recovery occurs within a community, and it includes both individual and contextual factors.

At the core of these ecological features is recovery capital, which draws on the social capital literature (Bourdieu, 1985), and involves the sum of personal and social resources that facilitate the process of recovery (Best & Laudet, 2010). This conceptualization provides a qualitative mechanism compatible with SAMHSA’s (2012) definition of recovery. Many studies have assessed the importance of individual markers of recovery capital (e.g., quality of life, hope, self-esteem, stress, wages) for sustaining recovery from substance addiction. Quality of life, a subjective measure of well-being, predicts abstinence after treatment over time (Laudet, Becker, & White, 2009). Additionally, hope influences motivation and behavior, and is characterized by three dimensions of agency, perceived and available pathways to achieve one’s goals (Snyder et al., 1991), and perceived and available opportunities in one’s environmental context (Stevens, Buchannan, Ferrari, Jason, & Ram, 2014). Higher levels of hope are associated with greater time abstinent and quality of life (Irving, Seidner, Burling, Pagliarini, & Robbins-Sisco, 1998). Self-esteem, a person’s self-reflection of their worth and abilities (Tafarodi & Milne, 2002), is negatively related to substance use, such that individuals with low self-esteem are more likely to engage in drug use compared to individuals with higher self-esteem (Jones & Heaven, 1998). Recovery home residents with higher levels of stability (i.e., agreeableness, conscientiousness, and low neuroticism) tend to have higher scores on self-esteem and lower levels of perceived stress (Reilly, Stevens, & Jason, 2017). Additionally, stress and self-efficacy, as well as social embeddedness, have been associated with quality of life (Jason, Stevens, Doogan, & Light, 2020). Finally, earning capital, like wages, is also important for recovery, by providing stability and expanding individual’s options for pursuing recovery (Cloud & Granfield, 2008).

Socially contextualized recovery, as measured through social recovery communities, also recognizes the importance of social capital for recovery. Some recovering individuals live within sober community living homes or participate in mutual self-help groups, like Alcoholics Anonymous (AA). Association with abstinent communities like these may facilitate recovery through the cultural capital such association provides, which can include mechanisms of social control (e.g., bonding, cohesion, and monitoring) and social learning (e.g., observation and imitation of behavior; Moos, 2007). Social support, which can consist of the provision of information, emotional guidance, positive appraisal, and tangible resources (e.g., financial assistance; Cohen & McKay, 1984), predicts abstinence (Longabaugh, Wirtz, Zywiak, & O’malley, 2010) and greater abstinence self-efficacy (Stevens, Jason, Ram, & Light, 2015). Recovery home residents who mark other house residents as highly important people to them are likely to be sober one year later (Jason, Davis, Ferrari, & Anderson, 2007). Those who endorse relationships with other house residents are more likely to remain abstinent, even two years later (Jason, Stevens, Ferrari, Thompson, & Legler, 2012). Additionally, when AA members, increase interactions with abstinent members, participate in AA, and have high abstinence self-efficacy, these factors together have a large influence on recovery (Kelly, Hoeppner, Stout, & Pagano, 2012). A recovery home resident’s hopefulness and sense of community can predict their quality of life at their recovery home (Stevens, Guerrero, Green, & Jason, 2018). Thus, research at these settings demonstrates the benefits of social capital as manifested by abstinent individuals in one’s day-to-day social interaction network, as abstinent communities promote social norms and values that motivate individuals to refrain from substance use. The bonds that residents build within these settings serve as motivation to engage in pro-social behaviors and refrain from destructive behaviors that can lead an individual to relapse (Polcin, 2009).

The level of recovery social capital within a setting might influence how long residents remain within those settings. This might be related to the congruence between individuals and the settings in which they belong. Within the recovery home literature, Beasley and Jason (2015) found affective experiences such as feeling like one fits with a community may influence engagement and disengagement. Therefore, the fit of an individual within a recovery house may influence how long they stay in these settings, and those who remain in these recovery settings for longer periods of time have increased chances of maintaining abstinence (Jason et al., 2007). A few recovery home studies have begun to examine these types of individual- and group-level dynamics (Jason, Light, Stevens, & Beers, 2014). For example, individuals who have more trouble fitting in socially due to social anxiety are more likely to leave a supportive communal living environment and subsequently relapse (Boddapati, Hunter, Jason, & Ferrari, 2014). Future studies addressing the nature of these individual- and group-level dynamics in more detail would help to more concretely identify how different elements of recovery capital are accessed and how such access might be maximized in recovery home environments.

From the studies above, it is clear that social capital indicators of recovery can occur at the level of individuals and their primary social environments (the aggregate characteristics of their residences with regard to these same measures). Thus, it provides a useful conceptual view of “recovery,” as a characteristic of individuals and their ecological social environments. Both individual- and group-level characteristics can independently affect outcomes (e.g., Snijders & Bosker, 2012). Although the research literature on recovery from substance dependence routinely reports strong links between recovery trajectories and one’s social environment (Vaillant, 1995), it is much less common for measured characteristics of these environments to be explicitly studied. However, it is at least conceivable for social capital to be measured by independent reports of the same environments from others sharing the environment, that is, other residents of the same home. In other words, it is possible for both individuals and their social environments (recovery homes) to be assigned values on some set of measures, and these indicators can have different predictive effects (e.g., relapse) at each of these individual and recovery home levels.

In the current study, we investigated the construct of social capital from the perspective of both recovery home residents and for the homes themselves. The idea of measuring the “level of recovery” present in a social environment is a novel extension of the broadly accepted idea that social environments have a major effect on the recovery success rates of the individuals within them. Past studies addressing recovery as a measurement construct have not taken “environmental recovery” into account, which can affect conclusions drawn about the individual-level construct (Gonzales, Hernandez, Douglas, & Yu, 2015; Reise, Ventura, Nuechterlein, & Kim, 2005; Sterling, Slusher, & Weinstein, 2008).

The current investigation consists of a multi-level confirmatory factor analysis (CFA) of a single, broad recovery capital factor at both the individual and house levels. For many important research questions, there are substantial advantages of a single broad factor over a multitude of lower level factors starting with but not limited to parsimony provided that such a single factor model provides a reasonable fit to the data. And just because a single factor provides a reasonable representation of recovery for some purposes does not mean that more nuanced and multifactorial representations of recovery would not be better for other purposes. Thus, our first goal was to explore the adequacy of a single broad factor to represent recovery at both the individual and environmental level using the tools of multi-level CFA. Our second goal, contingent on success in the first goal, was to use the latent factors at both levels to predict one of the most fundamental measures of success in recovery, that is, rate of relapse. We hypothesized that both the individual- and house-level recovery factors would have significant and distinct relationships to relapse.

Method

Participants

Jason, Stevens, et al. (2020), Jason, Wiedbusch, Bobak, and Taullahu (2020) estimate that there are over 270,000 individuals living in recovery homes in the United States. The current study was set in Oxford House (OH) recovery homes, a network of over 2000 self-governed recovery homes in the US, housing over 20,000 individuals. There are no professional staff in these houses, which are rented and gender-segregated, and which house from 6 to 12 individuals in recovery. Residents are required to follow three rules, including paying their fair share of the rent (usually from $100 to $125 per week), following the rules that help maintain the home, and abstaining from substance use. Olson and Jason (2011) found OH residents perceive recovery as including growth, accountability, surrendering to a higher power, empowerment, social support, self-worth, gratitude, diversity, structure, and sobriety, and thus a conceptualization compatible with SAMHSA’s (2012).

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 description of it from a project-provided script; 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 42.¹

Participants were part of a longitudinal study that collected information every four months over 2 years. Participants were recruited and interviewed by field research staff in face-to-face meetings. Participants were compensated $20 for completing each assessment. Permission to do this study was obtained by the DePaul University Institutional Review Board.

Measures

Wages for the last 30 days were computed by taking the square root of wages to reduce positive skewness; this was then was used as a continuous variable.

World Health Organization Quality of Life Assessment-Brief (Quality of Life; 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 whole measure for our sample was .89.

The Drug Taking Confidence Questionnaire (Self-Efficacy; Sklar, Annis, & Turner, 1999) is an 8-item survey that measures self-efficacy in terms of abstinence. Participants are asked to consider eight theoretical high-risk situations and rate how confident they would be of resisting the urge to use a substance given the hypothetical circumstances. For our sample, this measure had good reliability (α = .95).

The Rosenberg’s Self-Esteem Scale (SES; 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.” Items include “I think I have a number of good qualities,” “I take a positive attitude toward myself,” and “I feel I do not have much to be proud of” The internal reliability for our sample of the SES scale was .92.

The Perceived Stress Scale (PSS; Cohen, Kamarck, & Mermelstein, 1983) was utilized to measure the degree to which situations in participants’ lives are appraised as stressful. PSS consists of four items measured on a 5-point Likert scale ranging from “never” to “very often.” 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 PSS scale for our sample was .73.

The Interpersonal Support Evaluation List (ISEL; Cohen & Wills, 1985) measured three types of actual or perceived social support (tangible, appraisal, belonging). Tangible support refers to instrumental aid which refers to 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 ISEL consists of 12 items measured on a 4-point likert scale ranging from “definitely false” to “definitely true.” The internal reliability of the ISEL scale was .88 for our sample.

The Psychological Sense of Community (SOC; Jason, Stevens & Ram, 2015) is a 9-item scale utilized to measure participants’ sense of community. 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 SOC scale was used as a whole measure, and for our sample, the α = .91.

Snyder’s State Hope Scale (Hope; Snyder et al., 1996) was utilized to measure participants’ current state of hope. The Hope measure contains two subscales: 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.

Procedure and Sample

Residents of participating homes were able to enter the study at any point during 2 years, and data were collected every 4 months over seven waves. Participation was determined based on several criteria related to the process of censoring (a feature of survival models). First, participants had to have provided at least one wave of follow-up data beyond a first baseline interview; otherwise, their reason for leaving could not be determined and they could not have contributed any information to the survival modeling. Ultimately, our survival analysis used the individual- and OH-level latent factors as predictors of resident and OH-average rates of relapse.

For each wave of data collection, participants were classified as either evicted from the OH due to using alcohol or drugs, or not (left for some other reason or still in residence). With seven survey waves, after the baseline wave, a maximum of six follow-up time intervals were thus available for observing and predicting potential relapse. When the timing of event occurrence is only known to be in a relatively large interval of time (i.e., between one wave and the next), the analysis is known as a discrete time survival analysis. Of course, participants entering OHs after wave 1 had fewer follow-up intervals available than those in the study at wave 1, but the analysis included all such participants having at least one follow-up interval—that is, at least one period after their first study participation for which we could determine definitively whether they had left the house due to substance use or not.

There were 714 residents of the OHs during this period of time, of which 666 (93%) agreed to participate in this investigation. We excluded 64 of these participants from the current study.² Both the CFA and survival models used the same set of 602 individuals (out of 666 total participants, or 90%, henceforth referred to as the initial risk set). The amount of missing indicator data was minimal (96% of subjects had complete data) but in line with standard practice, the 23 subjects with partial data on the indicators were still included in all analyses.

Statistical Analyses

A single-level CFA operates on a correlation matrix for a set of measures thought to be indicators of underlying latent factor(s). The underlying latent factors are responsible for the correlations which represent “shared” variance. Multi-level CFA is possible because higher level structure (e.g., an OH) can have a net impact on both residents and the influences of residents on each other and this shows up in a separate correlation matrix for resident-level measures at each level, resident and house. The separate correlation matrices could lead to similar or radically different factor structure at each level, and it is multi-level CFA that allows us to test hypotheses about these potential differences. The implications for interpretation are that inappropriate single-level CFA models on data that is inherently multi-level may or may not be misleading and the only way to tell is to use the appropriate multi-level model. The significance of this approach is that clarity is gained on the nature of influence at the different levels of the hierarchy, which should lead to better programs for recovery.

For the CFA models, only the first baseline interview with each of the 602 analysis sample participants was utilized. For the combined measurement-plus-survival modeling, the same 602 individuals were included, but their data over time were used to define the relapse event. Survival models are distinct in part because of the way they handle the “survival variable,” that is, the binary event variable, which is 0 until the event occurs (in this study, a relapse event), 1 for the time period in which the event occurs, and then missing thereafter.^3,4

Internal validity of the recovery factor was assessed with a CFA using Maximum Likelihood Robust (MLR, robust with respect to non-normality of latent variables) conducted using Mplus 8.3 (Muthén & Muthén, 2017). Specific internal validity research questions included: does a single factor model adequately fit the data at the individual and the house level (what is referred to as 2-level)? If it does, is the latent factor variance significant at one, the other or both levels? The following variables were used as indicators: wages, self-efficacy, stress, self-esteem, social support, sense of community, hope, and quality of life. To reduce skewness and avoid potential problems with outliers, we transformed all but two indicators, quality of life and stress. We reversed the scaling of indicators with negative skewness (hope, self-esteem, self-efficacy, psychological sense of community, social support) so that they were positively skewed (similar to wages) and tried a variety of power transformations (e.g., square root) and picked the one that best reduced the positive skewness and resulted in a reasonably normal distribution. We fit the latent recovery factor model using each resident’s first survey in their first OH during the study period.

To address external (predictive) validity, we conducted a 2-level discrete time survival analysis in Mplus 8.3 (Muthén & Muthén, 2017) using the individual- and OH-level latent factors as predictors of resident and OH-average rates of relapse, respectively. Such an analysis is essentially a 2-level regression model (subject event histories at level 1 nested within an OH at level 2) for an observed binary outcome. However, the estimated outcome is the (probit transformed) hazard of relapse (the conditional probability of relapse given that it has not happened yet) at each of the k = 1, 2, … 6 discrete follow-up periods.

Preliminary models included quadratic and linear effects of follow-up time (to allow for changes in the hazard independent of predictors) but we found that only the linear term was necessary. Results are reported on the hazard probability scale rather than the probit transformed scale used for estimation, as the latter is more difficult to interpret. Both MLR and Bayesian estimation were used for the CFA models, giving nearly identical results, but for CFA, MLR results are reported, because a larger and perhaps more familiar set of fit statistics are available for such models. For the combined CFA-survival modeling, however, we used Bayesian estimation because it is the only computationally feasible approach and we are presenting the results in a way that is familiar to frequentists (point estimates and 95% CI’s, see Muthen & Asparouhov, 2012). However, fit statistics are largely non-overlapping, and so those presented differ for the CFA versus the CFA-relapse model results.

We used the Mplus default specification for a predictor with effects at both levels where the level 1 and 2 effects are both estimated independently of each other and each is represented by a latent variable. This corresponds to the so-called “group-mean” centering approach in the HLM literature (e.g., Raudenbush & Bryk, 2002) but with the important exception that rather than approximating latent variables with observed group means (for level 2 effects) and observed deviations from group means (for level 1 effects), Mplus uses the latent variable equivalents to reduce attenuation bias of the between level effect (Asparouhov & Muthén, 2019). In the context of our model, the predictor starts out as a latent variable defined by observed indicators but the approach is the same, we estimate within and between effects of the latent recovery factor separately.

Results

Descriptive Statistics

A total of 142 (24%) relapse events were observed during the study; 320 (53%) individuals exited without a relapse, 134 (22%) remained in the house at their last follow-up and 6 (1%) had a missing exit type. The analysis sample of 602 was 51% male, with a mean age of 37.0 years (SD = 10.5). Participants identified as European American (78.8%), African American (8.5%), Latinx (10.0%), with all other ethnicities accounting for 2.7% (Asian American, Alaskan Native, American Indian, and Pacific Islander).

Descriptive statistics for the recovery indices are shown in Table 1. For the most part, the transformed indicators are reasonably normally and continuously distributed (skewness and kurtosis are minimal) with the exception of self-efficacy, which before reverse scaling and transformation was quite high in our sample with 41% of the sample scoring at the maximum possible, a substantial ceiling effect. The reverse scaling and transformation of self-efficacy did not fix this and merely turned the ceiling effect in to a floor effect. As is the default, however, all Mplus MLR models are estimated using a correction factor to make the standard errors more robust to departures from normality so the floor effect is unlikely to distort the results.

Table 1.

Descriptive statistics for recovery indicators

Indicator	Mean	SD	Skew	Kurt	Min	Max	Min%	Max%	N
Stress	2.54	0.76	−0.02	−0.22	1.00	5.00	4.65	0.17	602
Quality of life	6.96	1.21	−0.28	0.40	2.54	10.00	0.17	0.17	600
Wagesb	27.46	16.78	−0.17	0.15	0.00	109.54	13.25	0.17	581
Self-efficacya	0.56	0.53	0.34	−1.14	0.00	1.71	41.03	1.33	602
Hopea	1.11	0.55	−0.38	−0.19	0.00	2.58	9.63	0.17	602
Self-esteema	0.76	0.43	−0.66	−0.72	0.00	1.67	18.60	0.17	602
Social supporta	0.69	0.44	−0.25	−1.01	0.00	1.63	18.94	0.17	602
Sense of communitya	0.61	0.52	0.36	−0.50	0.00	2.24	31.56	0.50	602

Kurt is kurtosis, a measure of non-normality. Min% and Max% are the percentages of the sample at the observed minimum and maximum scores.

^aIndicators have been reverse scaled and then root transformed to reduce skewness without inflating the variance or kurtosis.

^bWages is the sqrt(wages)/10.

Confirmatory Factor Analysis (CFA)

We found one latent factor fit both the individual and house levels. Our 1-factor CFA model fit reasonably well by standard fit indices not directly dependent on sample size (RMSEA = .040, CFI = .959, TLI = .942) but did not fit by the chi-square fit statistic, X² (40, N = 602) = 78.44, p < .001. It is not unusual for a reasonably well-fitting model to have a significant χ² fit statistic when the number of observations is considerably more than the degrees of freedom (Hu & Bentler, 1995). Inspection of modification indices indicated that several correlations among residual influences may be significant but given the quite acceptable RMSEA, we decided not to further modify the model.

Table 2 and Fig. 1 present the standardized indicator loadings and raw factor variances. All factor loadings and factor variances at both individual and OH levels were significant (p < .001). At the individual level, there were four strong indicators with standardized loadings that ranged in absolute value from .59 to .71 (self-esteem, hope, quality of life, and stress) and four weaker indicators with standardized loadings that ranged in absolute value from .18 to .49 (social support, self-efficacy, sense of community and wages). At the OH level, however, all indicators were strong with standardized factor loadings that ranged in absolute value from .58 to .99. In fact, for 4 (hope, social support, quality of life, and wages) of eight indicators, the OH-level residual variances (not shown in Table 2) were very small, near zero, and not significant. This is noteworthy because, in line with standard recommendations for Bayesian models (Asparouhov & Muthen, 2010), we set these residuals to zero when we combined the CFA model with the relapse model.

Table 2.

Confirmatory Factor Analysis (CFA) standardized factor loadings and raw factor variances

Variable	Effect	Level	Est.a	SE
Recovery factor	Variance	House	0.222	0.064
Hopeb	Loading	House	0.814	0.141
Social supportb	Loading	House	0.919	0.085
Sense of communityb	Loading	House	0.577	0.165
Self-esteemb	Loading	House	0.764	0.095
Quality of Life	Loading	House	−0.977	0.037
Stress	Loading	House	0.877	0.075
Self-efficacyb	Loading	House	0.681	0.194
Wages	Loading	House	−0.987	0.056
Recovery factor	Variance	Individual	0.567	0.101
Hopeb	Loading	Individual	0.691	0.040
Social supportb	Loading	Individual	0.487	0.048
Sense of communityb	Loading	Individual	0.259	0.049
Self-esteemb	Loading	Individual	0.709	0.026
Quality of Life	Loading	Individual	−0.673	0.041
Stress	Loading	Individual	0.594	0.038
Self-efficacyb	Loading	Individual	0.385	0.039
Wages	Loading	Individual	−0.182	0.049

^aAll estimates are significant at p < .001.

^bIndicators have been reverse scaled and then root transformed to reduce skewness without inflating the variance or kurtosis.

Fig. 1.

Standardized 2-level confirmatory factor analysis (CFA) model (measurements at respondent’s first observation)

Comparing factor loadings at the resident (individual) versus OH (house) levels, several indicators shifted rank order appreciably, the most notable being wages that went from the weakest resident-level indicator to the strongest OH-level indicator. Thus, the empirical indicators all contributed to the latent factor’s value both statistically and substantively at the resident level, the house level, or both. The total factor variance partition was 72% at the resident level and 28% at the OH level.

Relapse Survival Analysis

For a model with relapse as an outcome but without the latent recovery factor, MLR is computationally feasible so MLR results are presented. The estimated variance of the OH-level relapse latent variable on the probit scale is 0.140, which is highly significant using a nested chi-square test, X² (1, N = 602) = 13.29, p < .001. A complete and incisive summary of these results can be obtained by comparing a high relapse rate OH (95th percentile) in our sample versus a low relapse rate OH (5th percentile) in our sample on the hazard probability scale across follow-up time as shown in the two hazard curves in Fig. 2. The relative risks of a relapse between these two model fitted scenarios is substantial (0.270 vs. 0.072 at follow-up 1, about a 3.8-fold difference). In summary, relapse happens sooner and thus is more likely in some houses compared to others and these OH (house)-level differences are substantial.

Fig. 2.

Fitted hazard curves comparing a high sample relapse rate OH (95th) versus a low sample relapse rate OH (5th)

Combined CFA-Relapse Survival Analysis

We ran preliminary analyses with demographic variables (age, race/ethnicity, sex, education level), and as none of these variables were related to relapse, either singly or in combination, they were not included in subsequent analyses. The path model diagram for this analysis is shown in Fig. 3 and, as discussed above, this model was fit using Bayesian estimation because MLR is intractable. In line with standard recommendations, Fig. 3 and Table 3 provide Bayesian point estimates (medians of posterior distributions) with statistical significance in the frequentist sense reported on the basis of Bayesian credibility intervals (the counterpart to MLR confidence intervals) that do not contain zero. Similar to the 2-level CFA, the Bayesian posterior predictive p-value (the rough equivalent to the ML model chi-square p-value) was significant indicating a lack of fit. This could be due to correlations among residual influences in the CFA part of the model or specific effects of recovery indicators on relapse net of the recovery factor. It is beyond the scope of this work, however, to pursue these issues here but they represent interesting avenues for future research using Bayesian methods outlined in Muthen and Asparouhov (2012).

Fig. 3.

Standardized 2-level confirmatory factor analysis (CFA) model (measurements at respondent’s first observation)

Table 3.

Selected results from the combined 2-level CFA-relapse model

Type	Effect	Level	Est.	lo (2.5%)	hi (97.5%)	Sig
Raw	Relapse on recovery factor	Individual	0.124	−0.027	0.283
Raw	Relapse on time	Individual	−0.295	−0.422	−0.180	*
Raw	Recovery factor variance	Individual	0.616	0.517	0.726	*
Raw	Relapse on recovery factor	House	0.598	0.213	1.051	*
Raw	Relapse threshold	House	2.072	1.891	2.281	*
Raw	Recovery factor variance	House	0.230	0.131	0.411	*
Raw	Relapse residual variance	House	0.112	0.022	0.328	*
Std	Relapse on recovery factor	Individual	0.088	−0.019	0.199
Std	Pseudo-R-squared	Individual	0.160	0.084	0.260	*
Std	Relapse on recovery factor	House	0.660	0.255	0.912	*
Std	Pseudo-R-squared	House	0.435	0.065	0.832	*

Std indicates standardized parameter. Time is the linear effect of follow-up interval, 1–6. In the Sig column,

^*indicates significance in the sense that the Bayesian 95% credibility interval does not include 0.

The new information in the combined model is whether or not the latent recovery factor at the resident and OH levels predict resident and OH-level relapse. The effect of the resident (individual)-level latent recovery factor on relapse was small and not significant (standardized effect = .09, pseudo-R-squared = .16). On the other hand, the OH (house)-level latent recovery factor was a significant and strong predictor of relapse (standardized effect = .66, pseudo-R squared = .44). On the hazard probability scale at the first follow-up, comparing a resident high on the recovery factor to a resident low on the recovery factor, the relative risk of a relapse increases by a factor of about 1.7. The same comparison at the OH level results in the relative risk of a relapse increasing by a factor of about 4.5.

Discussion

This study examined eight measured indicators and found that the 1-factor model had a good fit (RMSEA = .04). External validation was evaluated by augmenting the CFA to predict relapse hazard for both individuals and house averages. This model found strong evidence of a hypothesized effect of OH-level recovery means on house-average relapse rates, but once this effect was accounted for, there was little additional evidence for an individual-level effect.

The recovery factor was conceptually sound, as all the indicators are well accepted, observed measures of recovery with face validity, internal and external validity, and reasonable reliability in our sample. In addition, they were inter-correlated in the way that would be theoretically expected. These inter-correlations make it possible to conceptualize a single higher order latent factor that captures the shared variance among the different facets. The utility and value of such higher order factors are well established in the social sciences. On the other hand, the existence of a reasonable higher order factor does not limit the use of lower order factors for more nuanced and detailed hypotheses.

The chi-square goodness of fit statistic for 1 factor suggests that there is room for improvement, and correlated residuals that might be significant if they were included in the model are a sign of additional factors. These additional factors could be measurement artifacts or they could be substantively important additional dimensions of recovery. The single factor had significant variance at both the resident and house level, which is a different issue than the goodness of fit of the 1-factor model, and all the indicators loaded significantly on the single factor.

We believe that there is value in examining this as a single factor versus examining each of these variables as a distinct level 1 and level 2 predictor of recovery. The main value is parsimony and the ability to entertain more complex analyses because fewer latent dimensions are involved, and the multi-level survival analysis demonstrates this. This would be difficult using eight lower order latent variables each with both individual- and house-level components. In addition, we are now engaged in additional lines of research involving complex network analyses that would not be feasible without a reasonably good, single, higher order recovery factor.

The findings of an individual- and house-level one-factor structure suggested by the CFA are somewhat more complicated than most CFA analyses. The individual-level latent factor represents a tendency for the measured indicators to correlate within individuals, that is, a low self-esteem rating is likely to be accompanied by low hope, low quality of life, and high stress. We may hypothesize that when an individual is successful at maintaining employment, this would be associated with increases in self-esteem, hope, self-efficacy, and less stress. Also, we found evidence that the house-level averages of the indicators also correlate within OHs; that is, a home with an average rating of low self-esteem among its residents is also more likely to have a low average rating for hope, quality of life, etc. It may be possible that some OHs are simply more effective than others at collectively promoting residents’ recovery; that is, in influencing such a recovery positively. Such differences could include the baseline composition of the home, availability of local community supports, presence of a few residents with many years of sobriety and considerable experience in how to maintain it, or a particular set of personalities who mesh particularly well or poorly.

The differential importance of wages for the individual-level versus OH-level recovery factors is also notable. At the individual level, wage was the lowest-loading indicator, less than a third as important (in terms of standardized factor loading) as the most important (highest loading) indicators such as self-esteem and hope. On the other hand, at the OH (house)-level, wages had the highest loading. It is plausible that one reason wage is such a strong house-level indicator of recovery is that contacts for possible employment—a major issue in this population—can be readily shared within OHs and thus constitute an important class of social capital.

The OH-level latent recovery factor is a strong predictor of average rates of relapse across homes. However, the individual-level factor did not predict individual variation in relapse rates. That is, in the current sample, the risk or protective aspects of “recovery” that affect the likelihood of relapse are features of the OHs, that is, their composition in terms of the latent recovery factor, and not of the individuals. In practical terms, this means, for example, that an individual’s probability of relapse in any given time period is primarily related to the average of the “recoveries” of his or her OH colleagues, and not of his or her own personal “recovery” status. An individual with a promising recovery profile in a home where most others’ profiles are more negative may not benefit from his or her state of recovery but could be negatively affected by the other less-recovered residents.

The specific implication of this pattern of findings for the population we studied is that recovery is a “team sport” and that OH residents stand or fall based upon the total recovery capital available in their home. Their state of recovery plays into the process only insofar as it contributes to the overall recovery capital available. Although the idea that the social environment is an important influence on recovery is certainly not new, the study’s results indicate it may be even more important than has been generally recognized.

These results could have implications beyond the OH population. At the very least, they are consistent with the idea that recovery can be strongly influenced by others with whom the recovering individual has regular social contact, whether for good or for ill. But the same questions are raised for the broader recovering population: How does one find and maintain recovery-positive social interdependencies, especially early in recovery when one has relatively little recovery capital to offer others? OHs achieve this, to the extent they do, through normative guidelines: the more recovered are supposed to help the less recovered. Evidence suggests that this norm is generally adhered to (Doogan, Light, Stevens, & Jason, 2019); however, even so, the current study suggests that some OHs are quite a lot more recovery-positive than others. Thus for both OH and non-OH environments, it would appear that the nature of the recovery-supportive relationships one develops need to be scrutinized, to better understand how they ultimately manage or fail to provide access to recovery-supportive social capital.

There have been a number of efforts to conceptualize recovery, thus tapping social capital domains. For example, Cale, Deane, Kelly, and Lyons (2015) used a scale developed for individuals with mental illness and substance use disorders, the Recovery Assessment Scale, but it might not be appropriate to generalize a scale from those with mental disorders, reflecting more severe problems, to those with substance use disorders, who are less likely to seek support and have less severe levels of dependence. Butler, Budman, Mcgee, Davis, Cornelli, and Morey (2006) developed the Addiction Severity Assessment Tool, which is used to assess the clinical outcomes of adult clients in substance abuse treatment. Of the six factors from this 27-item multi-dimensional self-report measure, only one measures recovery skills/self-efficacy and those five items assess a lack of beliefs and skills that could facilitate staying clean in the presence of stress. Matto, Miller, and Spera (2006) developed the Ecological Assessment of Substance Abuse Experiences questionnaire that measures the influence of family, peers, and community on clients’ attitudes and outcome beliefs related to substance abuse and recovery. Its underlying factors included items loading on “belonging-recovery,” “belonging-drugs/alcohol,” “disconnect-recovery,” and “attitudinal congruence-recovery.” Sterling et al. (2008) used items from multiple instruments that tapped recovery capital, and 23 items were observed loading on eight recovery capital factors; reliance on God and faith as a means of coping, sense of spirituality, recent sobriety, stable employment/income, alcohol/drug-free living environment, proportion of one’s life spent free from the effects of alcohol, satisfaction with both one’s marital situation/living arrangements, and levels of education/training.

Of the studies mentioned in the above paragraph, only Cale et al. (2015) reported correlations among the subfactors (and the correlations were suggestive of a higher order factor as the correlations were all significant at p < .001). Our study differs from prior work in two important ways including: (a) we explicitly tested the goodness of fit of our higher order single factor model and (b) the model testing explicitly reflected the multi-level context of residents nested within recovery homes. Of course, a single, higher order factor represents correlations or shared variance among lower order factors, and this is why higher order factors are useful and why total scores are often computed even when factor analysis clearly demonstrates correlated subfactors. But perhaps of greater interest was that the multi-level, single factor model fit the data well and in particular, the OH-level recovery factor was the stronger predictor in our data even though it is made up of individual-level measures. It is one thing to ask individuals about the quality of their recovery environment (e.g., an OH) or to get some recovery environment measures (e.g., OH characteristics) and use those in a standard single-level prediction model, but we believe it is an improvement to employ a multi-level model and isolate an environmental component that is strongly predictive. We believe substance use disorder researchers would benefit from focusing on how a group of interconnected individuals construct recovery environments that either promote or undermine their joint recovery. This cannot be accomplished with a single-level design even if the sampling is done correctly.

Imagine a carefully crafted, single-level design where we had 600 residents of recovery homes but we only sample a single resident in 600 different houses widely separated all across the United States. This does away with multi-level modeling because now we know that each resident is completely unrelated and independent of all the others and each resident has nothing to do with the recovery environment of any other resident in the study. In such a design, there is no house-level component to the recovery factor and so we would probably drastically underestimate the predictive power of the recovery factor if we are missing the house-level component. More importantly, we might misattribute whatever predictive power we find to individuals rather than the environment. Based on such research, we might recommend individual therapy to try to enhance an individual’s recovery factor and completely ignore their recovery environment likely resulting in a poor success rate.

In addition, higher order factors have their uses (independent of the multi-level issue) but when a research team like Sterling et al. (2008) develop an instrument that has eight factors, using so many factors may be cumbersome, time consuming, and at times overwhelming. That’s where a single, higher order factor is attractive, provided the lower order factors are substantially correlated. Our study, therefore, demonstrated that a single higher order factor works well with our particular eight measures, allowing us to make progress on other fronts, especially network modeling.

Regarding possible practical applications of this research, with an established recovery factor, such as the one that was identified in this article, it would be possible to examine recovery score changes in individuals and houses over time. It would then be possible to hypothesize what conditions might increase such recovery scores, particularly if they are associated with better outcomes over time. As an example, we could investigate whether having a close friend and individuals who are willing to lend resources to the resident improves recovery scores for those newly entering the houses. In addition, using network analyses, it would be possible to simulate housing conditions, such as settings with facilitating conditions (such as friendship and loans) and determine if those conditions led to improved recovery scores over time; these types of analyses are now been conducted by the investigators. Clearly, identifying the facilitating characteristics that promote recovery in these community settings could have enormous benefits to many recovery and treatment settings.

There are several methodological limitations in this study. First, the data may not be representative of other recovery populations. Also, those completing the measures were from only OHs in three states, and it is unclear whether these findings would generalize to other areas or other types of recovery homes. Finally, although 42 recovery homes seem like a substantial level-2 sample, and indeed, the level-2 recovery factor variance accounted for a substantial 28% of the total variance and the associated p-value was <.001, the relatively small number of residents within homes along with their somewhat surprising heterogeneity in terms of the measures used in this study (the factor indicators as well as relapse rates) suggests our sample is probably underpowered to detect more subtle but potentially important aspects of the recovery construct under investigation (e.g., subgroup effects or synergistic interaction effects involving the recovery factor to name just two). This observation implies that future studies would do well to try to find tractable ways to recruit larger samples of homes.

In summary, our study contributes to the literature by defining a domain of recovery, by combining several factors that represent individual and group-level elements of recovery capital. These types of findings further our understanding of complex individual and environmental domains. Both, but distinctly, contribute to recovery capital of those with substance use disorders. These issues have particular relevance to the field of Community Psychology, which values an ecological analysis that attempts to understand behavior in the context of individuals and communities (Kelly, 2006). Such work includes a better understanding of individuals’ and groups’ behavior in bidirectional interaction with their social contexts. This field values and appreciates the shaping influence of context, and the present study illustrates its powerful effects even over individual factors.

Highlights.

• House level factors predicted relapse. Individual level factors did not predict relapse.

• Residents status in recovery homes is based upon the total recovery capital available in homes.

• Recovery is strongly influenced by others with whom the recovering person has regular social contact.

• Individual’s probability of relapse was primarily related to the “recoveries” of home peers.

• Recovery capital available in homes highlights the importance of the social environment for recovery.