Domestic Violence Survivors’ Housing Stability, Safety and Well-being Over Time: Examining the Role of Domestic Violence Housing First, Social Support and Material Hardship

Abstract

Intimate partner violence remains a significant public health issue and survivors often need various forms of support to achieve safety. The increased likelihood of experiencing housing instability and homelessness among survivors has led to an uptake in domestic violence agencies implementing housing-based interventions, such as Domestic Violence Housing First (DVHF), to address survivors’ needs. The current study expands on prior research supporting the effectiveness of DVHF to examine situational factors that moderate the outcomes associated with this model among 406 survivors seeking services from domestic violence agencies located in the Pacific Northwestern region of the United States. Using latent profile analysis, participants were grouped into three latent classes: (1) “High Abuse/Instability,” (2) “Still Affected,” and (3) “Doing Better.” Latent transition analysis was used to estimate the probability that participants would transition into a different latent class over time with social support, material hardship, and receipt of DVHF services included as model predictors. Receipt of DVHF predicted improvements in survivors’ safety, housing stability, mental health, and well-being, such that receiving DVHF was associated with higher odds of survivors transitioning into the “Doing Better” class. Social support and material hardship also emerged as significant factors predicting class membership, such that higher levels of social support and financial stability predicted membership in the “Doing Better” class. Additionally, social support and financial stability appeared to augment receipt of DVHF services, with DVHF being more strongly associated with positive outcomes among participants who also had high levels of social support and lower levels of material hardship.

Survivors of intimate partner violence (IPV) who seek services from domestic violence (DV) victim service agencies come from all walks of life and circumstances (Davies & Lyon, 2013; Sullivan & Virden, 2017). Some survivors may be seeking minimal supports while others require extensive and possibly long-term help. The goal of these agencies is to provide the individualized services needed to ensure that all survivors are safer, more stable, and experiencing greater well-being after receipt of services (Cattaneo et al., 2020; Davies & Lyon, 2013; Sullivan, 2018). Avoiding homelessness and attaining affordable, safe housing is a tremendous hurdle for many IPV survivors exiting an abusive relationship (Chiaramonte et al., 2021; Pavao et al., 2007). With affordable housing becoming less available across the United States, agencies have accelerated their focus on helping unstably housed survivors obtain safe and stable housing.

One model that is being increasingly used by DV agencies to enhance survivors’ safety and housing stability over time is Domestic Violence Housing First (DVHF), a community-based intervention that entails providing unstably housed IPV survivors with a set of services that includes housing-focused advocacy and/or cash assistance (Sullivan & Olsen, 2016; Thomas et al., 2021). The goal of the intervention is for survivors to achieve safe and stable housing and psychological well-being. In a longitudinal demonstration evaluation comparing survivors who received DVHF with those who received services as usual (SAU), the DVHF model was more effective than SAU in improving survivors’ housing stability, safety, and mental health over a period of 24-months (Sullivan et al., under review). With growing evidence of the effectiveness of the DVHF model, it is now critical to examine the situational factors that impact the effectiveness of the intervention. Prior studies have confirmed that both social support and material hardship impact well-being, with social support having a positive effect (Voth Schrag et al., 2020) and material hardship having a negative effect (Tucker-Seeley et al., 2013). Therefore, the current study explored how the DVHF model, social support, and material hardship collectively predict survivors’ experiences of housing instability, abuse, mental health symptoms, and psychological well-being.

IPV, Mental Health and Well-being

IPV is a pervasive issue associated with a wide range of severe and long-lasting mental health consequences. Prior research has established a clear connection between IPV victimization and increased risk of mental health issues such as depression (Ahmadabadi et al., 2020; Beydoun et al., 2012; White & Satyen, 2015), anxiety (Chandan et al., 2020; Schaefer et al., 2021), and post-traumatic stress symptoms (Flemling et al., 2012; Signorelli et al., 2020). IPV has also been found to be negatively related to measures of subjective well-being among survivors, including their quality of life (Beeble et al., 2010; Cardenas et al. 2021; Wathen et al., 2016). However, DV support services can combat many of these negative outcomes. For example, cross-sectional research (Sullivan & Virden, 2017) suggests DV services may contribute to increases in hope, an indicator of well-being related to greater life satisfaction (Gilman et al., 2012; Munoz et al., 2017; Slezackova & Krafft, 2016; Wu, 2011). Though increasing survivors’ sense of hope and general psychological well-being is certainly a goal of DV agencies (Sullivan, 2018), the connection between IPV, psychosocial well-being and what leads to changes in survivors’ hopefulness over time is relatively understudied.

IPV and Material Hardship

Material hardship, defined as the inability to pay for essentials such as food, medicine, or housing, is distinct from the experience of income-based poverty, such that individuals can struggle with paying such bills even if they are living above the poverty threshold (Iceland & Bauman, 2007; Neckerman et al., 2016). Given that material hardship may lead to diminished social and emotional well-being (Black et al., 2009; Tucker-Seeley et al., 2013), identifying ways to mitigate material hardship is imperative.

IPV victimization can result in greater material hardship through several pathways. Abusive partners can block survivors from working or going to school (Adams et al., 2013; Showalter, 2016), steal their money or destroy their credit (Adams et al., 2020), or stalk them as they move from one residence to another (Baker et al., 2010; Chan et al., 2021; Desmond, 2016). The relationship between IPV and material hardship may also be more indirect. For example, IPV often leads to depression and PTSD (Ahmadabadi et al., 2020; White & Satyen, 2015), which can impact material hardship if survivors are psychologically unable to work or pay their bills (Voth Schrag, 2015).

Two large, longitudinal studies have established a strong link between IPV victimization and material hardship (O’Connor & Nepomnyaschy, 2020; Voth Schrag, 2015). Using data from the longitudinal Fragile Families and Child Well-being Study, O’Connor and Nepomnyashy (2020) confirmed that IPV is associated with greater material hardship, even after controlling for prior material hardship. Similarly, Voth Schrag’s (2015) national sample of over 2,700 women found that IPV correlated with greater material hardship, and this association was partially mediated by depression, signifying IPV may impact material hardship directly as well as indirectly through mental health.

Social Support, IPV, and Material Hardship

There is longstanding and robust evidence that social support is related to higher social and emotional well-being across numerous populations and circumstances (e.g., Siedlecki et al., 2014; Voth Schrag et al., 2020). Social support also decreases the odds of experiencing material hardship (Henley et al., 2005; Livermore et al., 2015), and is associated with shorter episodes of homelessness (Caton et al., 2005) as well as exiting homelessness (Zlotnick et al., 2003). It is not surprising, then, that higher social support has been found to increase IPV survivors’ safety and well-being (Beeble et al., 2009; Goodman et al., 2005; Ogbe et al., 2020).

Some IPV survivors lack social support as a result of the abusive (ex)partner intentionally isolating them from their family and friends (Myhill & Hohl, 2019). Others do not reach out for social support due to fear for their friends’ and family’s safety (Goodman & Smyth, 2011), and still others do not feel that they have supportive networks to reach out to (Trotter & Allen, 2009). Regardless of the precipitating factors, inadequate social support results in less emotional and material support provided by the survivors’ social networks (Voth Schrag et al., 2018).

Higher social support, on the other hand, has been shown be a protective factor for IPV survivors (Stylianou et al., 2021) For example, Beeble and colleagues (2009) reported that IPV survivors who experienced both high levels of abuse and low levels of support had more negative mental health outcomes over time than did other survivors. Conversely, higher social support has been shown to protect survivors from re-abuse and to improve their well-being (Bybee & Sullivan, 2005). Through both emotional and instrumental support, friends and family can provide IPV survivors with encouragement, information, protection, and access to resources that lead to greater safety and well-being over time.

Current Study

The aim of the current study was to examine the associations among DVHF, social support, material hardship, and survivors’ psychological well-being. The primary goal of DVHF is that survivors will be safer, stably housed and experience improved well-being. The importance of these separate but related factors, which may or may not move in concert with one another, led to our grouping participants based on measures of housing instability, safety, mental health symptomatology, and quality of life using latent profile analysis (LPA). We then used a longitudinal extension of LPA, latent transition analysis (LTA), to assess the likelihood of participants transitioning into a different latent class over time and to identify factors that predicted those transitions.

Method

Data were derived from a larger, longitudinal study examining the impact of the DVHF model on DV survivors’ safety, housing, and well-being. The parent study utilized a quasi-experimental, naturalistic design intended to typify the real-world experience of IPV survivors receiving DV services (see Sullivan et al., 2021 for additional details). The design involved inviting all unstably housed IPV survivors to participate in the study within two weeks of their approaching a DV agency for help, and then interviewing them every six months over a 24-month period (baseline, 6mo, 12mo, 18mo, 24mo). The design allowed us to capture the typical variation in services offered by DV agencies, which is tied to the resources available to them (e.g., whether they have flexible funding available; staff turnover impacting advocacy services). A total of 406 participants were recruited from five DV agencies in a Pacific Northwest state of the United States that offered housing-inclusive advocacy and flexible funding as it was available. Participants were paid $50 for each interview, and the Michigan State University Institutional Review Board (IRB) granted approval of the study. For the current study, we analyzed data from the baseline and 6-month follow-up interviews (hereafter referred to as Time 1 and Time 2).

Measures

Structured interviews included a number of standardized measures asked at each time point. The following measures from the larger study were included in the analyses:

Services received.

The six-month interview included questions about services received, including whether participants had received shelter, transitional housing, support groups, counseling, advocacy, and referrals. They were also asked if staff helped them “work on housing and getting other things” they needed from the community in the prior six months.

Flexible funding received.

Each agency documented the financial assistance provided to participants. Entries included dates, amounts, and descriptions of how funds were used.

Safety.

The Composite Abuse Scale (CAS; Loxton et al., 2013) was slightly modified to measure physical, emotional, and sexual abuse, as well as stalking over the prior six months. Specifically, the items “harass you at work” and “hang around outside your house” were replaced with “repeatedly follow you, phone you, and/or show up at your house/work/other place” and four questions were added (“stalk you,” “strangle you,” “force sexual activity,” and “demand sex whether you wanted to or not”). The response options for the current study ranged from 0 - 5: 0 “never,” 1 = “once,” 2 = “several times or between 2-3x in the last 6 months,” 3 = “once a month,” 4 = “once a week,” and 5 = “daily.” Four subscales measured physical abuse (alpha = .91), emotional abuse (alpha = .91), sexual abuse (alpha = .92), and stalking/harassment (alpha = .84). Mean scores were computed and Cronbach’s alpha for the 31-item scale was .95.

The 14-item Revised Scale of Economic Abuse (SEA2; Adams et al., 2020) assessed abusive tactics targeted toward damaging intimate partners’ and ex-partners’ economic stability over the prior six months (e.g., does your partner “keep you from having a job or going to work,” “keep financial information from you”). Response options ranged from 0 = never to 4 = quite often and mean scores were computed (Cronbach’s alpha = .91).

Housing instability.

The seven-item Housing Instability Scale (HIS), created specifically for this study, included six items from the 10-item The Housing Instability Index (Rollins et al., 2012) as well as: “In the last 6 months, have you been homeless or had to live with family or friends to avoid being homeless?” Scores range from 0-7, with higher scores indicating greater instability. The scale demonstrates strong predictive and concurrent validity, as well as scalar equivalence over time (Farero et al., 2022). Coefficient alphas were examined at each wave of data collection and Cronbach’s alpha for the study was .79.

Material hardship.

Material hardship was measured by the 2-item Inability to Make Ends Meet subscale from Barrera et al.’s (2001) Scale of Economic Hardship. The scale refers to financial difficulty experienced over the prior 6-months (3 months in the original scale). Response options for “how difficult has it been to pay your bills in full?” ranged from 0 “not at all difficult” to 3 “very difficult.” Response options for “Having money left over at the end of the month” ranged from 1 “more than enough money left” to 5 “very short of money.” Scores were summed, with higher scores indicating greater material hardship.

Depression.

Depression was assessed by the 9-item Patient Health Questionnaire (PHQ-9) (Kroenke, Spitzer, & Williams, 2001). Responses ranged from 0 “not at all” to 3 “nearly every day,” and referred to feelings over the prior two weeks (e.g., “little interest or pleasure in doing things”). Sum scores were computed, and Cronbach’s alpha was .88.

Anxiety.

The 7-item Generalized Anxiety Disorder scale (GAD-7; Spitzer et al., 2006) was used to assess anxiety. Items included “not being able to stop or control worrying” and “worrying too much about different things.” Responses were in reference to individuals’ feelings over the prior two weeks using a scale that ranged from 0 “not at all” to 3 “nearly every day.” Sum scores were computed, and Cronbach’s alpha was .91.

Post-traumatic stress disorder (PTSD) symptomatology.

The 10-item Trauma Screening Questionnaire (TSQ) assessed for symptoms of PTSD (Brewin et al., 2002). Participants were asked to think about the abuse they had experienced and to indicate (no = 0, yes = 1) if they had experienced any of the listed PTSD symptoms (e.g., “feeling upset by reminders of the event”) at least twice in the prior week. Scores could range from 0 to 10; a score of 6 or higher indicates probable PTSD (Cronbach’s alpha = .75).

Well-being.

Well-being was assessed through measures of quality of life and hope. Quality of life was measured by a 9-item scale adapted from Andrews and Withey (1976). Survivors were asked how satisfied they felt about parts of their lives over the prior 6-months (e.g., “How do you feel about your life overall?”) Responses ranged from 1 “terrible” to 7 “extremely happy.” Mean scores were computed, and Cronbach’s alpha was .88.

The 12-item Herth Hope Index (Herth, 1992) was used to measure hope. Each item was associated with either positive or negative outlooks on the survivor’s current situation (e.g., “I have a positive outlook toward life”). Response options ranged from 1 “strongly disagree” to 4 “strongly agree.” Mean scores were computed and Cronbach’s alpha was .71.

Analyses

Retention was 92% (n = 375) at the six-month follow-up. Participants retained in the study were compared to those not retained with regard to age, race, ethnicity, number of children, history of homelessness, history of abuse, relationship status, and mental health symptomatology and no statistically significant group differences were found.

Determining who received DVHF or SAU.

Agency data and survivor interviews were used at the six-month time point to categorize survivors as having received DVHF or SAU. Thirty participants had received no services and were therefore not assigned a value for this variable. Of the 345 participants who did receive services between Time 1 and Time 2, participants who had received housing-focused advocacy and/or funding were categorized as having received DVHF (64%, n=221) (see Sullivan et al., 2023 for more detail). The other participants (36%; n=124) were considered to have received SAU because they had received DV services but: 1) had not received housing-focused advocacy, despite needing such help; and 2) had not received cash assistance.

Latent profile analysis.

LPA is a person-centered analytic approach, meaning that it seeks to identify patterns or profiles that describe subgroups within a larger population (Magnusson, 1998). These subgroups—also called latent classes—are identified through participants’ responses to a set of indicator variables. The goal of the analysis is to explain heterogeneity in the data by identifying latent classes in which participants are more similar to others in their latent class than they are to the sample as a whole. To determine the configuration of latent classes that best represent the data, LPA solutions are iteratively compared such that each model is evaluated against a model with one fewer latent class, with the goal of finding the most parsimonious model (i.e., the model with the fewest number of latent classes) that adequately represents the data. In the current study, LPAs were estimated independently at Time 1 and Time 2, such that measures of housing instability, safety, PTSD, anxiety, depression, quality of life, and hopefulness collected at baseline formed the latent classes at Time 1, and those same measures collected at the six-month interviews formed the latent classes at Time 2. All variables were centered using the global mean (e.g., average of scores from Time 1 and Time 2 together) in order to make the relative differences between classes more easily interpretable while maintaining information about mean-level changes over time (Moeller, 2015).

To select the best fitting model, we considered both statistical and conceptual indicators of fit (Collins & Lanza, 2010). Statistical indicators of model fit included the Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), and sample-size adjusted Bayesian Information Criterion (aBIC), with lower values indicating better model fit. Likelihood ratio tests, including the Lo-Mendell-Rubin (LMR) test and the Bootstrapped Likelihood Ratio Test (BLRT), were used to identify a point at which adding additional classes no longer improved model fit. The substantive meaning of each latent class solution was evaluated to identify a model that fit the data statistically and meaningfully informed the topic of study better than the other candidate models (Masyn, 2013).

Latent transition analysis.

Latent transition analysis is a longitudinal extension of LPA in which latent class solutions are modeled at each time point and the probability of transitioning between classes over time is estimated. After identifying the best fitting latent class solution at each time point, we carried out longitudinal measurement invariance testing to evaluate the conceptual and statistical equivalence of the solutions across time. Statistical indicators will usually preference freely estimated models (i.e., models with no measurement invariance parameters imposed), as free estimation can capture small nuances in the data. However, measurement non-invariance can pose challenges for analysis and interpretation, and the methods literature recommends that measurement non-invariance parameters only be used when the difference between the latent class solutions at each time point are substantively meaningful in addition to statistically supported (Collins & Lanza, 2010). Measurement invariance testing was approached through this lens.

Next, to estimate the probability of remaining in the same or transitioning into a different latent class over time, we fit an unconditional LTA (i.e., an LTA that did not include auxiliary variables). Finally, we added predictors into the model and ran a conditional LTA to estimate the ability of DVHF, social support, and material hardship to predict class membership and class transition. The conditional LTA was carried out using the 3-step method, which incorporates auxiliary variables in such a way that the latent classes remain defined only by the indicator variables (Asparouhov & Muthén, 2013). Time-varying measures of social support and material hardship were incorporated as predictors at Time 1 and Time 2. Receipt of DVHF services was added to the model as a predictor at Time 2 only, as services had not yet been received at Time 1. Racial minority identity and foster care history were also tested as time-invariant predictors, however they were dropped given their non-significant associations with Time 1 class membership and model constraints (i.e., retaining non-significant categorical variables would have restricted the generalizability of the transition probability estimates).

All LPA and LTA analyses took place in Mplus version 8.6 (Muthén & Muthén, 1998-2021) using the Maximum Likelihood (MLR) estimator, which is robust to non-normality. For the LPA, missing data was addressed with the Full Information Maximum Likelihood estimation, which keeps participants who lack complete data in the model for latent class estimation. Conditional LTA models include only participants who had full data on all predictor variables (Asparouhov & Muthén, 2013), which reduced the sample size from n = 406 to n = 339. We compared participants who were included in the conditional LTA (n = 339) to those who were not (n = 67) and found no significant differences between the two groups on any LCA indicator variable nor any demographic variable with sufficient variability to support an analysis (identifying as a racial or ethnic minority, participant age, citizenship status, and employment within the last six months). This indicates that the full sample was comparable to the reduced sample used for the conditional LTA.

Results

Latent Profile Analysis – Can survivors be grouped based on varied experiences of housing instability, safety, mental health, and well-being?

To best reflect survivors’ varied experiences of housing instability, abuse, mental health, and well-being, two through five class solutions were computed separately for the Time 1 and Time 2 data. Model fit statistics (presented in Table 1) were evaluated in the context of model interpretability. At Time 1, information criterion indicators continued to decrease with each successive class (i.e., a global minimum was not reached before identification problems began to occur). The BLRT also remained significant in each successive solution, but the LMR LRT was significant through the three-class model only, suggesting that the three-class model offered the best statistical fit. The three-class model was substantively meaningful, as well. Class 3 (24% of participants) endorsed higher rates of housing instability and abuse than either of the other two latent classes, high rates of anxiety, depression, and PTSD, and low quality of life and hopefulness, and was therefore named the High Abuse/Instability class. Class 1 (35% of participants) reported rates of housing instability and abuse that hovered around the global mean but that were substantially lower than the High Abuse/Instability class. In addition to these relatively lower rates of abuse, Class 1 endorsed notably lower rates of anxiety, depression, and PTSD, and reported greater quality of life and hopefulness. This class was therefore named the Doing Better class. Class 2 (42% of participants) was a unique mix of the previously discussed classes, with rates of housing instability and abuse quite similar to the Doing Better class but mental health and well-being indicators closely resembling the High Abuse/Instability class. Based on these patterns, this class was named the Still Affected class. Table 2 presents participants’ socio-demographics by latent class, and descriptive data on the latent class indicators are presented overall and by latent class in Table 3.

Table 1.

Model Fit Indicators for 2 – 5 Class LPA Solutions at Time 1 and Time 2

Assessment Point	No. of Classes	AIC	BIC	aBIC	LMR	BLRT	Entropy
Time 1 (baseline)	2	16640.52	16776.73	16668.85	< 0.01	< 0.01	0.90
	3	16168.75	16353.04	16207.07	< 0.01	< 0.01	0.87
	4	15925.54	16157.91	15973.86	0.09	< 0.01	0.89
	5	15776.90	16057.34	15835.22	0.38	< 0.01	0.88
Time 2 (6 months)	2	14842.33	14928.73	14858.93	0.25	< 0.01	0.97
	3	13724.17	13857.68	13749.81	0.05	< 0.01	0.91
	4	13091.71	13272.35	13126.41	0.52	< 0.01	0.92
	5	12671.11	12898.87	12714.85	0.37	< 0.01	0.94

Note. AIC = Akaike Information Criterion, BIC = Bayesian Information Criterion, aBIC = Sample-Size Adjusted Information Criterion, LMR = Lo-Mendell Rubin Adjusted Likelihood Ratio Test, BLRT = Bootstrapped Likelihood Ratio Test

Table 2.

Socio-Demographics of Sample by LCA Class, at Baseline

	High Abuse/Instability (n = 98)	Still Affected (n = 169)	Doing Better (n = 139)	Total (n = 406)
	M (SD)	M (SD)	M (SD)	M (SD)
Age	34 (8.11)	36 (9.67)	33 (8.54)	35 (9.02)
	% (n)	% (n)	% (n)	% (n)
Gender
Female	96.9 (95)	95.9 (162)	97.8 (136)	97 (393)
Male	1.0 (1)	3.0 (5)	2.2 (3)	2.2 (9)
Genderqueer/non-conforming	2.0 (2)	1.2 (2)	0 (0)	1.0 (4)
Race/Ethnicity (choose all that apply)a
Non-Hispanic White only	37.8 (37)	40.2 (68)	28.3 (39)	35.6 (144)
Hispanic/Latinx	29.6 (29)	31.4 (53)	43.2 (60)	35.6 (142)
Black	17.3 (17)	17.2 (29)	21.6 (30)	18.7 (76)
Asian	5.1 (5)	5.9 (10)	0.7 (1)	3.9 (16)
US Indigenous	16.3 (16)	10.1 (17)	10.8 (15)	11.8 (48)
Middle Eastern	1.0 (1)	2.4 (4)	0 (0)	1.2 (5)
Any minoritized racial/ethnic identity	62.6 (61)	59.8 (101)	71.1 (99)	64.4 (261)
Primary language English	84.7 (83)	78.7 (133)	77.7 (108)	79.8 (324)
Heterosexual	82.7 (81)	85.2 (144)	89.9 (125)	86.2 (350)
Parenting minor children*	68.4 (67)	69.2 (117)	82.7 (115)	73.6 (299)
Education (at baseline)
Less than high school	23.5 (23)	26.6 (45)	35.3 (49)	28.8 (117)
High school graduate/GED	27.5 (27)	21.9 (37)	18.0 (25)	21.9 (89)
Vocational training/some college	23.5 (23)	39.1 (66)	34.5 (48)	36.2 (147)
Bachelor’s degree	11.2 (11)	7.7 (13)	7.9 (11)	8.6 (35)
Advanced degree	4.1 (4)	4.7 (8)	4.3 (6)	4.4 (18)
U.S. citizen	84.7 (83)	81.1 (137)	79.9 (111)	81.5 (331)
Prior history of homelessness	82.7 (81)	71.0 (120)	69.1 (96)	73.2 (297)
Homeless as a child	30.6 (30)	22.5 (38)	20.9 (29)	23.9 (97)
Foster care	22.5 (22)	15.4 (26)	15.1 (21)	17.0 (69)

^*Significant difference at p < 0.05

^aBecause participants could choose multiple categories, significance testing was conducted based on the any minoritized racial identity variable rather than individual race/ethnicity categories.

Table 3.

Study Variable by Full Sample and LCA Class at Time 1 and Time 2

Time 1
	High Abuse/Instability (n = 98)	Still Affected (n = 169)	Doing Better (n = 139)	Total (n = 406)
	M (SD)	M (SD)	M (SD)	M (SD)
Housing Instability	5.39 (1.42)	4.68 (1.63)	4.48 (1.77)	4.78 (1.67)
Physical Abuse	2.66 (1.02)	0.84 (0.62)	0.87 (0.73)	1.29 (1.09)
Emotional Abuse	3.56 (0.87)	1.74 (0.99)	1.43 (1.07)	2.07 (1.31)
Sexual Abuse	2.82 (1.52)	0.69 (1.04)	0.58 (1.09)	1.16 (1.51)
Economic Abuse	2.47 (0.90)	1.19 (0.82)	1.09 (0.94)	1.46 (1.05)
Stalking	4.00 (1.06)	1.80 (1.36)	1.56 (1.25)	2.25 (1.60)
PTSD	8.29 (1.71)	7.94 (1.45)	4.60 (2.34)	6.88 (2.48)
Depression	15.45 (4.76)	15.95 (3.62)	5.24 (3.22)	12.16 (6.28)
Anxiety	16.53 (6.03)	16.18 (4.89)	6.60 (3.99)	12.99 (6.73)
Quality of Life	3.61 (1.17)	3.60 (0.95)	4.85 (0.94)	4.03 (1.16)
Hopefulness	2.98 (0.53)	2.92 (0.47)	3.39 (0.38)	3.09 (0.51)
Social Support	3.09 (1.17)	3.12 (1.13)	3.60 (1.11)	3.28 (1.15)
Inability to Make Ends Meet	7.12 (1.32)	6.58 (1.62)	6.24 (1.74)	6.59 (1.63)
Time 2
	High Abuse/Instability (n = )	Still Affected (n = )	Doing Better (n = )	Total (n = 375)
	% (n)	% (n)	% (n)	% (n)
Received DVHFa	48.6 (18)	52.3 (79)	71.7 (124)	64.1%
	M (SD)	M (SD)	M (SD)	M (SD)
Housing Instability	5.00 (1.78)	3.75 (1.92)	2.83 (1.97)	3.41 (2.04)
Physical Abuse	1.68 (0.92	0.17 (0.29)	0.10 (0.25)	0.29 (0.60)
Emotional Abuse	2.69 (0.74)	0.49 (0.66)	0.27 (0.51)	0.60 (0.92)
Sexual Abuse	1.50 (2.17)	0.04 (0.18)	0.03 (0.22)	0.18 (0.65)
Economic Abuse	2.17 (0.96)	0.38 (0.64)	0.25 (0.55)	0.49 (0.84)
Stalking	3.38 (1.20)	1.20 (1.28)	0.58 (0.81)	1.11 (1.34)
PTSD	8.19 (1.68)	7.85 (1.80)	3.76 (2.67)	5.85 (3.07)
Depression	14.78 (4.11)	14.42 (4.01)	4.56 (3.32)	9.54 (6.19)
Anxiety	15.68 (4.60)	14.75 (5.38)	5.37 (3.76)	10.17 (6.61)
Quality of Life	3.48 (1.13)	3.89 (1.14)	5.26 (0.97)	4.54 (1.28)
Hopefulness	2.83 (0.52)	2.90 (0.49)	3.45 (0.41)	3.17 (0.53)
Social Support	2.62 (0.95)	3.19 (1.19)	3.89 (1.05)	3.48 (1.18)
Inability to Make Ends Meet	6.95 (1.35)	6.25 (1.76)	5.49 (1.93)	5.94 (1.87)

^aPercentages for the Received DVHF variable were calculated based on the n = 245 participants who received services between Time 1 and Time 2.

For the Time 2 data, information criterion indicators did not reach a global minimum and the BLRT remained significant with each successive model. Unlike at Time 1, the LMR LRT was not significant with any number of classes, though it did border significance for the three-class model (p = 0.05). Conceptually, the three-class model offered the best substantive fit, as well, producing meaningful latent classes that were well-distinguished from one another. The Time 2 three-class model was very similar to the three-class model at Time 1, with latent classes still classified as Doing Better (50%), Still Affected (40%) and High Abuse/Instability (10%). Due to the conceptual meaningfulness and similarity to the latent class model that was statistically supported at Time 1, the three-class model was selected as the final Time 2 model and moved forward for use in the LTA.

Latent Transition Analysis – Do survivors transition into different latent classes over time?

Measurement invariance testing.

As expected, measurement invariance testing indicated that the freely estimated model was a better statistical fit for the data (see Table 4). Evaluation of the substantive differences supported this conclusion, with the freely estimated model showing substantive improvement in all three latent classes from Time 1 to Time 2, such that across latent classes the intercepts were slightly lower for levels of abuse and mental health symptoms, and higher for indicators of well-being. This improvement was determined to be relevant to the research question, and the measurement component of the LTA therefore allowed for free estimation of model parameters. The freely estimated longitudinal models are presented in Figure 1.

Table 4.

Longitudinal Invariance Model Fit Indicators

	Free Parameters	TRD p-value	AIC	BIC	aBIC
Freely Estimated	92	< .001	29260.46	29629.04	29337.12
Fully Constrained	59	--	29726.75	29963.10	29775.90

Figure 1.

Model Mean Values for Freely Estimated Longitudinal Model (N = 406)

Size of latent classes in the unconditional longitudinal model.

The unconditional LTA was based on the full sample (N = 406). The proportion of participants assigned to each latent class in the LTA was similar to the proportion of participants assigned to each class in the LPAs. The Still Affected class remained the largest class at Time 1 (40%), followed by the Doing Better (36%) and the High Abuse/Instability (24%) classes. At Time 2, the Doing Better class was the largest (50%), followed by the Still Affected (41%) and the High Abuse/Instability (9%) classes.

Latent transition probabilities in the unconditional longitudinal model.

Estimated latent transition probabilities from the unconditional LTA are presented in Table 5. The probability of stability (i.e., staying in the same class over time) is bolded along the diagonal. These values indicate that participants who were assigned to the Doing Better class at Time 1 had an 88% probability of remaining in the Doing Better class at Time 2. Conversely, participants first assigned to the High Abuse/Instability class had only a 24% probability of remaining in that class at Time 2, with higher probabilities of transitioning into the Still Affected (46%) or Doing Better (30%) classes. Participants assigned to the Still Affected class at Time 1 had a 63% probability of remaining in that class at Time 2, with 9% transitioning into the High Abuse/Instability class and 28% transitioning into the Doing Better class. Collectively, these results indicate that participants’ housing instability, safety, mental health, and well-being tended to improve over time, though there was a substantial amount of variation in that trajectory, particularly for participants first assigned to the Still Affected class.

Table 5.

Unconditional Probability of Class Transition (N = 406)

Time 2
Time 1	High Abuse/Instability	Still Affected	Doing Better
High Abuse/ Instability	0.24	0.46	0.30
Still Affected	0.09	0.63	0.28
Doing Better	0.01	0.11	0.88

Size of latent classes in the conditional longitudinal model.

The conditional LTA included the n = 339 participants for whom there was data on receipt of DVHF services, social support, and material hardship. The proportion of participants assigned to each latent class in the conditional model was very similar to the proportions found in the unconditional model, with 24% of Time 1 participants assigned to the High Abuse/Instability class, 43% of participants assigned to the Still Affected class, and 33% assigned to the Doing Better class. At Time 2, 9% of participants were assigned to the High Abuse/Instability class, 40% were assigned to the Still Affected class and 51% assigned to the Doing Better class.

Ability of Material Hardship, Social Support, and DVHF to predict class membership.

The conditional LTA identified multiple significant relationships between predictor variables and the latent classes (see Table 6). Material hardship significantly predicted membership in the High Abuse/Instability class compared to the Doing Better class at both time points, such that greater material hardship predicted a significantly higher likelihood of being in the High Abuse/Instability class. Material hardship was less able to differentiate the Still Affected class from the Doing Better class, not reaching significance at Time 1 and with significant results but odds ratios indicating a potentially very small effect size at Time 2. Social support was significantly related to class membership at both time points, with higher social support predicting a significantly higher likelihood of being in the Doing Better class compared to the High Abuse/Instability or Still Affected classes. Receipt of the DVHF intervention compared to services as usual also significantly differentiated the High Abuse/Instability and Still Affected classes from the Doing Better class, such that receiving DVHF predicted a significantly higher likelihood of being in the Doing Better class.

Table 6.

Logistic Regression Coefficients for the Conditional LTA with the Doing Better Class as Reference Group (n = 339)

	Predictor	Odds Ratio [95% CI]
Baseline
High Abuse/Instability	Social Support	0.68 [0.51, 0.90]
	Ends Meet	1.54 [1.21, 1.96]
Still Affected	Social Support	0.64 [0.50, 0.82]
	Ends Meet	1.18 [0.99, 1.40]
Time 2 (6 mo.)
High Abuse/Instability	Intervention	0.25 [0.09, 0.67]
	Social Support	0.44 [0.30, 0.65]
	Ends Meet	1.52 [1.14, 2.02]
Still Affected	Intervention	0.31 [0.15, 0.67]
	Social Support	0.58 [0.42, 0.81]
	Ends Meet	1.22 [1.01, 1.48]

Note. Bolded cells are significant at a p < .05 level. Racial minority identity and history of foster care were also originally included as time-invariant predictors regressed onto the Time 1 latent class variable but were not significant. Because all predictors must be set at pre-specified levels to explore transition probabilities at different predictor levels, retaining non-significant categorical variables would have restricted the generalizability of the transition probability estimates. These variables were therefore not included in the conditional analysis.

Ability of Material Hardship, Social Support, and DVHF to predict class transition.

The predictor variables’ effects on transition probability odds ratios further illustrated the regression results. With the other predictors held constant, participants who received the DVHF intervention were over three times as likely to transition from the Still Affected class at Time 1 into the Doing Better class at Time 2 than were participants who received services as usual. The odds were over four times higher for participants first assigned to the High Abuse/Instability class (see Table 7).

Table 7.

Transition Probability Odds Ratios for Receipt of DVHF with Social Support and Material Hardship Held Constant

Time 2
Time 1	High Abuse/Instability	Still Affected	Doing Better
High Abuse/Instability	--	1.27 [0.54, 2.98]	4.09 [1.49, 11.23]
Still Affected	0.79 [0.34, 1.84]	--	3.22 [1.50, 6.91]
Doing Better	0.25 [0.09, 0.67]	0.31 [0.15, 0.67]	--

Note. Bolded cells are significant at a p < .05 level.

The LTA Calculator feature in Mplus was used to estimate latent class membership and class transition with predictor variables fixed at specified levels. With the goal of exploring the complex relationships between DVHF services, social support, and material hardship, models were estimated to show the probability of class membership and transition among people who received DVHF but had low (20^th percentile), medium (mean) and high (80^th percentile) levels of social support and ability to make ends meet¹. While previous results highlighted the impact of DVHF with other variables held constant, these estimates illustrated the substantially different outcomes one can expect among participants who received DVHF services depending on their levels of social support and material hardship (see Table 8). These estimates are illustrative but should be understood as modeling estimates only, as there may be a limited number of participants whose experiences reflect the combination of levels modeled.

Table 8.

Class Membership and Transition Probability Models with Social Support and Ability to Make Ends Meet at Low (20^th percentile), Medium (mean), and High (80^th percentile) Levels

	Received DVHF, SS and Ends Meet Low	Received DVHF, SS and Ends Meet @Mean	Received DVHF, SS and Ends Meet High
Class Membership Probabilities	Time 1 35% High Abuse/Instability 49% Still Affected 16% Doing Better	Time 1 24% High Abuse/Instability 45% Still Affected 22% Doing Better	Time 1 13% High Abuse/Instability 34% Still Affected 53% Doing Better
	Time 2 19% High Abuse/Instability 53% Still Affected 28% Doing Better	Time 2 6% High Abuse/Instability 37% Still Affected 58% Doing Better	Time 2 1% High Abuse/Instability 16% Still Affected 83% Doing Better
Transition Baseline to Time 2	High Abuse/Instability at T1 31% High Abuse/Instability at T2 51% Still Affected at T2 17% Doing Better at T2	High Abuse/Instability at T1 13% High Abuse/Instability at T2 44% Still Affected at T2 43% Doing Better at T2	High Abuse/Instability at T1 4% High Abuse/Instability at T2 26% Still Affected at T2 70% Doing Better at T2
	Still Affected at T1 15% High Abuse/Instability at T2 68% Still Affected at T2 18% Doing Better at T2	Still Affected at T1 6% High Abuse/Instability at T2 54% Still Affected at T2 40% Doing Better at T2	Still Affected at T1 2% High Abuse/Instability at T2 32% Still Affected at T2 66% Doing Better at T2
	Doing Better at T1 4% High Abuse/Instability at T2 15% Still Affected at T2 81% Doing Better at T2	Doing Better at T1 1% High Abuse/Instability at T2 6% Still Affected at T2 93% Doing Better at T2	Doing Better at T1 <1% High Abuse/Instability at T2 2% Still Affected at T2 98% Doing Better at T2

The probability of class membership at Time 2 was notably different depending on levels of social support and ability to make ends meet. Participants who received DVHF but had low social support and ability to make ends meet had only a 28% probability of being in the Doing Better class at Time 2, whereas participants who received DVHF and had high social support and ability to make ends meet had an 83% probability of being in the Doing Better class at Time 2. The probability of class transition in those different circumstances was striking, as well. DVHF recipients who were classified into the High Abuse/Instability class at Time 1 had a 17% probability of transitioning into the Doing Better class at Time 2 if they had low social support and ability to make ends meet, compared to a 70% probability of making that transition if they had high social support and ability to make ends meet. These estimated probabilities illustrate the degree to which DVHF services, social support, and material hardship jointly predict housing instability, abuse, mental health, and well-being among intimate partner violence survivors.

Discussion

To our knowledge, this study is the first to group IPV survivors by their similar patterns of experiences as a step toward understanding what factors may help them achieve combined safety, housing stability and well-being over time. Three latent classes emerged from the analyses, as did clear factors that predicted membership in those latent classes over time. It is encouraging that the High Abuse/Instability class, comprised of survivors reporting higher levels of abuse, higher levels of housing instability, and lower levels of mental health and well-being, was the smallest group at baseline (24%) and became even smaller six months later (9%). Equally encouraging is that the Doing Better class became the largest group at six months (from 36% to 50% of the sample). While the Still Affected class remained about the same size over time (40% at baseline; 41% at six months), those who transitioned out of that class were more likely to move into the Doing Better class than into High Abuse/Instability class. All participants were receiving DV services, and these findings therefore align with prior evidence of DV agencies’ positive impact on survivors’ lives (Goodman & Epstein, 2008; Sullivan & Goodman, 2019; Sullivan & Virden, 2017).

As hypothesized, receiving the DVHF model predicted positive changes for survivors regardless of their level of abuse, housing instability and well-being at study entry. Six months after participants entered services, receipt of DVHF predicted being in the Doing Better class compared to both the High Abuse/Instability and Still Affected classes. Receiving DVHF seems especially important for survivors in the High Abuse/Instability class at Time 1, as they were four times as likely to transition into the Doing Better class at Time 2 than were survivors in the High Abuse/Instability class who did not receive DVHF. If participants began in the Still Affected class at Time 1 and then received DVHF, they were three times as likely to transition into the Doing Better class at Time 2. This suggests that the model is effective across levels of abuse, housing instability, depression, anxiety, PTSD, quality of life and hopefulness.

Material hardship also predicted group membership and transition over time. Consistent with prior research demonstrating a link between IPV and material hardship (e.g., O’Connor & Nepomnyaschy, 2020; Postmus et al., 2012; Voth Schrag, 2015), survivors reporting lower levels of material hardship were more likely to be in the Doing Better class. Results also aligned with earlier evidence indicating that social support is a protective factor for IPV survivors (e.g., Beeble et al., 2009; Bybee & Sullivan, 2005; Stylianou et al., 2021). Having higher levels of social support predicted being in the Doing Better class six months after entering services, compared to both the High Abuse/Instability and Still Affected classes.

While DVHF, material hardship, and social support independently predicted group membership, there is also evidence that the impact of DVHF is moderated by survivors’ levels of social support and material hardship. If they started out in the High Abuse/Instability class at study entry, and then received DVHF while also having high social support and ability to make ends meet, survivors had a 70% probability of transitioning into the Doing Better class six months later. If, however, they had low social support and ability to make ends meet, they had only a 17% probability of moving into the Doing Better class. These findings speak to the importance of understanding survivors’ experiences with DV services within the larger context of their lives.

Limitations

Generalizability of findings should be considered in light of the socio-demographics of the study participants. By design, the study participants were unstably housed or homeless and had also sought assistance from a DV program. Further, while this was a racially and ethnically diverse sample, there were few US Indigenous, Asian, or Arab participants. Most survivors were cisgender, heterosexual women, and the majority were parents. Additionally, the absence of study participants who received no services limits our inferences regarding the overall efficacy of DV services. Future studies that include participants who have not sought services from DV agencies will allow us to ascertain which changes are due to services received, and which are due to the passage of time or other factors. Such studies could be particularly useful in identifying what individual or community-level assets could be bolstered in the absence of formal DV services to reduce survivors’ experiences of instability and abuse and to increase their mental and emotional well-being.

The study design was intentionally naturalistic in order to document the real-world experiences of IPV survivors receiving community-based services. The chosen methodology allowed us to capture the reality that some survivors receive flexible funding from DV agencies, for example, not due to their need but rather to the resources available to agencies over time. It is typical that funding fluctuates, as does staff availability and expertise (Sullivan et al., 2021), which then impacts service delivery. We also cannot rule out that there were additional factors that may have impacted who received DVHF and who received SAU. However, these likely reflect differences that would be seen in other implementations of DVHF. Similarly, while our analysis of potential differences between the samples used for the unconditional and conditional LTAs did not identify any significant differences, the samples could differ based on unidentified characteristics or experiences.

There are also limitations inherent to latent profile and latent transition analysis as methods. Latent profile analysis describes latent subgroups seen in the data but cannot prove that those subgroups exist (Vermunt & Magidson, 2002). Particularly in light of the mixed statistical support for the three-class LPA solutions, these results should be understood as a starting place, rather than conclusion, for this line of research. Use of the freely estimated longitudinal model requires recognition that the latent classes do not mean exactly the same thing at Time 1 and Time 2. However, the strong conceptual similarity between the latent class models at each time point supports discussion of class stability and transition, and the improvement across latent classes over time is a meaningful finding in and of itself. It is also worth noting that our study included a limited number of auxiliary variables to maximize generalizability of the transition probability estimates in the conditional model, and we hope that future research will explore additional factors that may moderate the strength of the relationships we observed in this study. Finally, it should be noted that these analyses capture only six months of much longer-term experiences. Because the number of possible transition patterns increases exponentially with each time period added, we chose to focus on the first six months only. Our conclusions should be interpreted in light of these limitations.

Research Implications

Service-seeking DV survivors come with a variety of needs upon intake. Research that helps DV staff and advocates efficiently and effectively determine the best course of support for each survivor is incredibly important to preventing homelessness and future DV. This study made an important step towards understanding how to better help survivors transition into lives reflected by safety, housing stability and wellbeing. It is noteworthy, for example, that the largest group at Time 1 was the Still Affected class, comprised of survivors whose recent experiences mirrored the Doing Better class (lower rates of abuse and housing instability) but whose psychological wellbeing was more similar to the High Abuse/Instability class. These were survivors who, despite lower abuse and housing instability compared to the High Abuse/Instability class, were still experiencing higher mental health symptoms and lower quality of life. The persistence of these emotional and psychological effects even at lower levels of abuse highlights the often long-term effects of DV and the importance of understanding the varied impacts it may have on survivors’ lives. The salience of social support for this latent class also emphasizes that social support is not only crucial when abuse and instability are at their worst, but in subsequent stages of recovery and rebuilding, as well. Future studies should focus more on this group of survivors with the goal of identifying other factors that could support their holistic well-being.

Policy and Practice Implications

Results highlighted that DVHF services were linked to improvements in survivors’ housing instability, experiences of abuse, mental health symptoms, and overall wellness. These findings will hopefully encourage practitioners to expand their services to include housing and financial assistance and, crucially, encourage policy makers to increase the funding available for these services. Additionally, the higher odds of being in the Doing Better group when participants reported higher levels of social support compared to both the High Abuse/Instability and Still Affected group suggest that social support is an important resource for domestic violence survivors. Many practitioners already provide programming focused on fostering such support, and these findings can be used to bolster funding applications to maintain or expand those services. Finally, the experiences represented by the Still Affected class may provide practitioners helpful language with which to normalize the experience of many survivors, namely that one’s subjective experience does not always move in tandem with objective circumstances and there may be a delay between a decrease in abuse and instability and improvements in mental health and wellbeing.

While evidence from this study is directly useful for community-based DV agencies, there is only so much they can do without community and societal supports. The high rates of housing instability and mental health concerns resulting from IPV victimization necessitate that practitioners and policymakers recognize and promote factors that facilitate survivors’ continued safety, stability, and well-being. Further, state and federal policy changes are needed to address the myriad of structural issues facing IPV survivors, including the expansion of affordable housing options, accessible social services, and policies focused on ameliorating social inequities (Keefe & Hahn, 2021).

In summary, this study expands on earlier evidence that providing survivors with housing-focused advocacy and/or cash assistance through the DVHF model is more effective than services-as-usual in helping survivors achieve safe and stable housing, and shows that the DVHF model works regardless of survivors’ levels of abuse, housing instability, and well-being when they enter services. Further, this study demonstrates the importance of strengthening survivors’ support networks and addressing the extent of their material hardship to improve their outcomes. These concrete findings should be used by policy makers and practitioners to reduce the re-victimization of survivors and enhance their stability and well-being over time.

Public Policy Relevance:

The current study expands on prior evidence that providing survivors of intimate partner violence with housing-focused advocacy and/or cash assistance may help them achieve safe and stable housing. Results confirmed that the model works regardless of survivors’ levels of abuse, housing instability, and well-being when they entered services, and demonstrated the importance of strengthening survivors’ support networks and addressing the extent of their material hardship. Policies supporting the widespread provision of this model could result in greater safety, stability, and well-being for a larger number of survivors.