More than 1.2 million people were incarcerated in US state or federal prisons in 2022 (Carson, 2023). The percentage of women in the prison system has grown from 4% of the population in the 1980s to 7% in 2022 (Carson, 2023). In 2022, 64.6% of women in federal prison were serving time for drug offenses compared to only 44.6% of men (Carson, 2023). Further, while nearly half of US prisoners met criteria for a substance use disorder (SUD), only 33 to 46% of prisoners received treatment (Maruschak et al., 2021). Individuals with SUD are more likely to reoffend and have multiple incarcerations, indicating that SUD treatment is critical for reducing recidivism (Baillargeon et al., 2010; Zgoba et al., 2020). Additionally, non-completion of diversion programs has been linked to higher reoffending rates (Ozturk et al., 2022). Conversely, completion of diversion programs has been linked not only with reduced likelihood of returning to prison (Clifasefi et al., 2017; Bernard et al., 2020; Mueller-Smith and Schnepel, 2021; Ozturk et al., 2022), but also reduced likelihood of substance use overdose (Bernard et al., 2020), reduced cost to state and local governments (Anglin et al., 2013), improved likelihood of being employed (Clifasefi et al., 2017; Mueller-Smith and Schnepel, 2021) and access to stable housed (Clifasefi et al., 2017) with effects lasting up to 20 years (Mueller-Smith and Schnepel, 2021).

In response to overall increases in drug-related incarcerations since the “War on Drugs” in the 1980s (Drug Abuse Office and Treatment Act of 1972, 1972; Bureau of Justice Statistics, 1995; Harvey et al., 2007; Carson, 2023), there are calls for improved access to SUD treatment and alternative programs to divert drug offenders from long-term incarceration (National Institute on Drug Abuse, NIOH, U.S. Department of Health and Human, and Services, 2020; Caulkins et al., 2021). SUD diversion programs often offer reduced or commuted sentences for the completion of a specific program or set of requirements (usually involving SUD treatment) set out for the offender. Outcomes for such programs have been the focus of several review papers, with the focus often being on rate of successful program completion (Hartford et al., 2006; Harvey et al., 2007; Loveland and Boyle, 2007; Sirotich, 2009; Brown R. T., 2010; Brorson et al., 2013; Shapiro et al., 2014; Bernard et al., 2020; Lindquist-Grantz et al., 2021). Rates of completion vary between studies, but reviews suggest dropout rates may be slightly higher for diversion programs at 40–50% (Harvey et al., 2007; Brown R., 2010; Shaffer, 2011) compared to an average of 30.4% across SUD treatment programs types (Lappan et al., 2020).

There is a growing literature examining predictors of successful SUD diversion program completion and recidivism. This literature commonly concentrates on demographic factors including gender (Gray and Saum, 2005; Zettler, 2019; Gallagher et al., 2020), race (Hartley and Phillips, 2001; Gray and Saum, 2005; Ho et al., 2018; Zettler, 2019), and age (Ho et al., 2018; Shannon et al., 2018; Zettler, 2019); and socioeconomic factors such as educational attainment (Hartley and Phillips, 2001; Butzin et al., 2002; Dannerbeck et al., 2006; Ho et al., 2018), employment (Hartley and Phillips, 2001; Butzin et al., 2002; Dannerbeck et al., 2006; Gallagher et al., 2015; Gill, 2016; Ho et al., 2018; Shannon et al., 2018; Zettler, 2019; Gallagher et al., 2020), marital status and social relationships (Dannerbeck et al., 2006; DeVall and Lanier, 2012; Ho et al., 2018), and criminal history (DeVall and Lanier, 2012; Ho et al., 2018; Shannon et al., 2018; Gallagher et al., 2020). These studies suggest that a younger age at time of matriculation, greater extent of criminal history, minority status, less than high school education, and unemployment, may relate to lower likelihood of SUD diversion program completion. However, there are numerous inconsistencies across studies, which may be due to variations in treatment programs and populations – but also likely speaks to the difficulty of predicting mental health and substance use related outcomes including treatment completion (Gowin et al., 2019; Symons et al., 2019).

Research focused on demographic factors predicting treatment outcome is valuable in helping to understand which individuals need additional or augmented treatment. However, there have been calls to shift focus to more malleable predictors and to better understand under what conditions dropout occurs, to help inform what to target with treatment modifications or augmentations (Zgoba et al., 2020). The current study extends this call to criminal diversion programs in an attempt to understand additional factors related to dropout and increase retention in these programs.

Research focusing on how comorbid mental health and psychological variables relate to program graduation may be particularly useful for identifying SUD treatment augmentation strategies. For instance, comorbid depression symptoms and/or diagnosis at matriculation has been linked to higher rates of attrition from SUD treatment (Curran et al., 2002; Gray and Saum, 2005; Evans et al., 2009; Hickert et al., 2009). Hickert et al. (2009) found that depression in the 30 days prior to treatment more than doubled the likelihood of treatment termination. Evidence for relationships between other comorbid conditions (e.g., anxiety, posttraumatic stress disorder, psychosis) is less clear, but generally supports increased dropout for individuals with dual diagnoses (Levin et al., 2004; Gray and Saum, 2005; Lejuez et al., 2008; Evans et al., 2009; Szafranski et al., 2017; Zettler, 2018). However, it is important to note that some studies suggest comorbid mental illness is not associated with termination from diversion or treatment programs (Brown R., 2010; Gallagher et al., 2015) or may increase retention (Amaro et al., 2007; Hesse, 2009; López-Goñi et al., 2021), which also may be due to variations in study populations and treatment programs. For full systematic reviews highlighting some of these inconsistencies see Brorson et al. (2013) and Lappan et al. (2020). Dimensional assessments of behavior or personality have also been explored in relation to treatment completion. Specifically, impulsivity, hostility, aggression, and sensation seeking have been identified as potential predictors of treatment dropout in substance use programs (Loree et al., 2015; Hershberger et al., 2017; Choate et al., 2021). However, one limitation of this previous literature is that it has focused on predominantly male samples. Choate et al. (2021) examined whether gender moderated the relationship between personality traits and treatment termination. They found gender moderated this relationship such that some personality traits (e.g., grandiosity and hostility) related to treatment termination only among men, whereas other traits (e.g., submissiveness) predicted outcomes for women. Further highlighting the need for representation of women in SUD literature is the suggestion that women may have specific needs and barriers to treatment, including lack of services for pregnant women and/or affordable childcare, fear of losing custody, greater transportation concerns, economic concerns, higher rates of comorbid psychiatric conditions, increased likelihood of trauma, greater social stigma and discrimination, and less social support (Greenfield et al., 2007; Tuchman, 2010). While there has been prior work examining gender as a potential predictor of outcome for substance use or criminal diversion programs, further work is warranted to examine factors predicting treatment outcomes and treatment completion for women specifically.

The aim of the current study was to use machine learning (ML) techniques to examine predictors of graduation from a criminal diversion program for women facing SUD related prison sentences. ML is well-suited to this aim because it has the capability to describe complex statistical relationships among high-dimensional data. In such cases, ML has the potential to make better predictions than traditional modeling techniques, improving medical prognoses and diagnostic accuracy (Obermeyer and Emanuel, 2016). The use of ML methods is growing in the SUD literature as a way to examine treatment predictors (Barenholtz et al., 2020); however, it has been less commonly used to examine diversion program completion (Delen et al., 2021). The current study included demographic variables as well as a large battery of psychological variables assessing domains of mental health symptoms, sleep, physical activity, personality, and cognitive functioning. Ultimately, understanding characteristics related to graduation for women enrolled in an SUD criminal diversion program has the potential for informing modification of interventions to increase the likelihood of program completion.

The goal of the current investigation was to use data driven techniques to examine predictors of graduation for women participating in an SUD diversion program. Previous work dedicated to identifying predictors of treatment completion often emphasizes the relation between demographic and socioeconomic variables and graduation. This study extends this literature through the inclusion of these variables as well as a large battery of psychological variables assessing domains of mental health symptoms sleep, physical activity, personality, and cognitive functioning. Further, the current study explored graduation in women, a population historically underrepresented in SUD literature. The larger goal was to understand characteristics associated with program completion, with the aspiration to identify strategies for possible intervention or modification of such programs. Unfortunately, even with 147 variables, the ML models were unable to provide individual-level prediction of graduation better than the base rate. However, several variables were identified as being related to graduation with medium to large effect sizes including employment/financial risk, housing safety, anger/hostility, antisocial friends, impulsivity, trauma history, and interoceptive awareness. Many of these identified variables could be potentially targeted in SUD treatment and diversion programs to help improve completion rates.

At the forefront of this discussion, it is important to recognize that the use of ML or other predictive modeling has the potential to perpetuate bias against minoritized populations by relying on factors such as socioeconomic background to make decisions about sentencing (e.g., those sentenced to prison versus criminal diversion programs; Klingele, 2015; DiBenedetto, 2019). We suggest that the identification of such predictors is most beneficial if used to identify potential avenues to enhance treatment retention or success for all individuals enrolled in diversion programs, rather than to identify which individuals to enroll in these programs. ML has been heralded as an approach that may be able to predict at the individual-level mental and substance use related outcomes including treatment completion (Acion et al., 2017; Mak et al., 2019); however, this study and others demonstrate the difficulty of predicting outcomes, even when there is a breadth of information available to include in the models (Gowin et al., 2019; Symons et al., 2019). This may suggest more fine-grained temporal data, such as ecological momentary assessment (Shiffman, 2009), is needed to better predict program completion at the individual level. While our current ML models failed to predict graduation better than chance, the variable importance and logistic regression analyses identified several variables that could potentially be targeted to increase completion rates and may be important to explore in future research. Thus, while the goal of being able to indicate the probability of success of program graduation for any one individual entering treatment remains elusive, the identification of cognitive or behavioral factors relating to treatment completion remain important for informing potential targets for treatment augmentation at the group level (Czajkowski et al., 2015).

In prior work, insight and recognition of one’s problems have been related to increased motivation for treatment, treatment attendance, and abstinence from substances (Raftery et al., 2020). Similarly, in the current study, the CTQ Denial subscale was related negatively to graduation, which may suggest that women who have more insight into their prior experiences of trauma or who are more willing to report trauma, may engage more meaningfully in treatment. Additionally, trauma-informed substance use programs or programs that treat SUD and trauma concurrently have a lower rate of treatment dropout compared to treatment as usual (Amaro et al., 2007; López-Goñi et al., 2021). The current study found trauma to be positively related to graduation, which is seemingly in contradiction to previous studies indicating trauma or trauma-related symptoms relate to worse treatment outcomes (Hyman et al., 2008; Szafranski et al., 2017). This may be related to fact that those who are high in denial are also going to report less trauma experiences and thus, reduced trauma history may instead reflect reduced awareness or willingness to report. However, this finding may also be due to the uniqueness of the current sample. First, the current study examined only women and previous studies recruited majority male samples. Secondly, the women in the current study were recruited from a trauma-informed diversion program. This may have impacted treatment seeking behaviors such that those with significant trauma history may have been more likely to seek out PTSD treatment during WIR. Taken together, findings suggest that insight into one’s mental health concerns and trauma-related care may be important to consider in reducing attrition for women-focused substance use and criminal diversion treatment programs.

The current study also found that UPPS-P Lack of Premeditation was negatively associated with program completion. The UPPS-P Lack of Premeditation subscale assesses one’s tendency to act impulsively without consideration of consequences. A recent meta-analysis found pre-treatment Lack of Premeditation to be a robust predictor of poor therapy outcomes in SUD (Hershberger et al., 2017). Further, this meta-analysis examined pre to post treatment changes in impulsivity and reported a lack of significant change for Lack of Premeditation, Positive Urgency, and Lack of Perseverance and only small effect sizes for the changes in Negative Urgency and Sensation Seeking (Hershberger et al., 2017). Together, these findings suggest that impulsivity is related to treatment completion and that standard substance use treatments may not effectively target impulsivity, indicating potential benefit in identifying strategies that may do so. Various novel treatment augmentation approaches, such as cognitive remediation (Anderson et al., 2021; Nardo et al., 2021) and episodic future thinking (Aonso-Diego et al., 2021; Forster et al., 2021) have shown promise in reducing impulsivity in substance use or other populations. Future research and funding are warranted to identify whether impulsivity interventions in SUD or diversion programs may show promise for better retention and treatment outcomes.

Interoceptive awareness was another psychological construct identified as related to WIR program completion. Increasingly, research has shown a connection between diminished interoceptive awareness and substance use disorder (Smith et al., 2020a; Stewart et al., 2020). The current study suggested that not distracting oneself from bodily sensations was related negatively to graduation with a medium effect size, specifically that more distracting increases likelihood of graduating, and interoceptive self-regulation also had a small positive effect size with respect to graduation (Cohen’s d = 0.37). Thus, greater awareness of and attempts to regulate interoceptive sensations (via distraction or self-regulation) may be positive indicators for SUD or criminal diversion program completion. Distraction can be both adaptive and maladaptive depending on the context (Wolgast and Lundh, 2017; Waugh et al., 2020), specifically distraction for avoidance is often considered maladaptive; however, distraction for emotion regulation can be adaptive and is considered a useful distress tolerance technique to regulate intense emotions (e.g., as used in dialectical behavioral therapy; Linehan, 2014). It may be that for individuals recovering from SUD, distraction from internal cues, such as cravings, is an adaptive strategy. Treatment strategies that may hold promise for targeting interoceptive awareness and regulation include general distress tolerance and emotion regulation strategies (Linehan, 2014) as well as mind–body interventions, such as mindfulness (Black and Amaro, 2019), yoga (Petker et al., 2021), or Mindful Awareness in Body-Oriented Therapy (Price and Hooven, 2018; Price et al., 2019a,b).

Additional factors which may be useful to address in treatment are employment/finance and housing safety risk factors, which were negatively associated with program completion. While this is the first study to examine the WRNA as it relates to a diversion program for substance-related offences, previous studies have found WRNA scales to be significantly, positively related to subsequent arrests, convictions, and failure to complete probation (Van Voorhis et al., 2010). Additionally, previous studies have found employment to be significantly related to diversion program completion (Dannerbeck et al., 2006; Ho et al., 2018; Shannon et al., 2018). These findings highlight the need to address residential and occupational concerns early in treatment, as doing so may lead to higher rates of program completion.

This study used a data-driven framework to identify variables contributing to drop out from a women’s substance use criminal diversion program. While the ML model failed to predict graduation rates at the individual level a rate higher than chance, examination of VI and results from logistic regression analyses identified several factors relating to treatment completion with medium to large effect sizes – including trauma, interoceptive awareness, and impulsivity. Future studies are warranted to examine whether interventions aimed at processing trauma or improving interoceptive awareness and impulsivity are related to increased retention in substance use treatment settings.