Recidivism

The HKT-R (Spreen et al., 2014 ) is a structured professional tool for assessing the risk of violent recidivism in forensic psychiatric patients. The HKT-R consists of 33 factors spread over three domains: 12 Historical, 14 Clinical, and seven Future factors. All factors are rated on a 5-point scale, ranging from 0 to 4, in which “0” represents no risk and “4” represents a high level of risk. The Historical domain relates to the offender’s personal history up to the moment of the arrest for the current index-offense (e.g., judicial history, employment history, and victim type). The Clinical domain contains 14 factors that are divided into seven risk (e.g., impulsivity and hostility) and seven protective factors (e.g., coping skills and cooperation with treatment). The Clinical domain refers to the offender’s behavior in the last 12 months (e.g., problem insight, psychotic symptoms, and antisocial behavior). In our study, all protective factors were recoded so that higher scores indicated higher protection against reoffending (“0” represents no protection and “4” represents high protection). The Future domain is related to the assessment of potential risks, which could emerge after discharge from the FPC (e.g., stressful circumstances, living arrangements, and work situation).
For patients with a TBS order, a risk assessment of the Clinical items must be performed at least once a year. The annual scores on the 14 Clinical factors indicate whether a reduction in risk factors and/or an improvement in protective factors has occurred, compared with the previous 12 months of stay in the institution. Hence, if changes occur, that could assumingly be ascribed to the given treatment. In this study, only the Clinical dynamic factors were included as Historical factors are static and irreversible, and Future factors are exclusively related to the situation after release. Internal consistency for the Clinical domain was good at both measurement points, with Cronbach’s α being α_T1 = .80 and α_T2 = .83, respectively. Descriptive statistics of the clinical risk and protective factors are presented in Table 1.

A 30-year-old incarcerated African American (AA) male with a history of an abdominal gunshot wound was admitted from jail with diarrhea and hypovolemia. Computed tomography (CT) showed terminal ileal inflammation. Colonoscopy was consistent with Crohn’s ileitis. Ustekinumab (UST) was selected as initial therapy and infused on the day of discharge. This choice was based on leveraging the weight-based initial dose and ability to ensure adherence, with confirmation of future dosing through nurse-led injection appointments, and consistent access to medication with patient assistance programs. Unfortunately, due to transportation issues, he had multiple missed doses with a gap of as long as 6 months. However, he was reloaded with an IV dose, and with intensive social work support to coordinate logistics, he received one maintenance dose prior to being released from his detention center. Unfortunately, he remains lost to follow-up by both the clinic and specialty pharmacy, a challenge not uncommon in relation to the correctional system and recidivism.

We will summarize baseline and demographic characteristics using means and standard deviations (or interquartile ranges and medians) for continuous variables and percentages for categorical variables. We will report the proportion of eligible clients consenting to participate and other descriptive data on outcomes such as treatment compliance rates.
For identifying the causal impacts, our main analysis within the RCT will focus on two estimands. First, we will analyze outcomes among the intent-to-treat (ITT) sample, or those offered access to ODR regardless of whether they use the tool or have an eligible case. Because unobserved features of clients—for instance, their tech-savviness or how pressing their legal issue is—impact how clients use the tool, this provides the strongest causal estimate of the tool’s effect. Second, we will analyze outcomes among the “opt in and eligible sample”(TOT), which represents the effect of the assistance + tool on those who actually used the tools to resolve legal issues.6 This treatment-on-treated estimand is relevant for policymakers’ interpretations of the value of ODR for substance users—it corresponds to the counterfactual, “what would happen if we adjust the estimates to assume that we are able to tailor the offers to those who will actually use the tools and who have eligible cases to use them for?”
Our estimation approach will depend on the specific outcome variable in question and will be specified in greater detail in a later pre-analysis plan. Broadly, for the intent-to-treat “ITT,” we will examine using non-parametric differences in means (since covariates, in expectation, will be balanced across the groups due to the randomization) and linear and logistic regression, where the coefficient of interest is one on a binary variable for offer of treatment [49 ]. For the treatment-on- treated “TOT,” we will use the two-stage least squares approach where we regress our measure of compliance (“uses the ODR”) on the offer of the legal tool, and then use the fitted values from step one to measure the impact on outcomes. We will also use survival models for time-dependent outcomes (e.g., time to relapse; time to recidivism; time to mortality).

The West Midlands Police and Hampshire Constabulary datasets differ in the set of available variables and therefore we analyze them separately to take advantage of the extra information where available, and to avoid data pooling assumptions. This study is unique in that the impact of the intervention is examined across two independent samples. It is also unique in that it uses two different crime severity measures.
The following measures were used in the analysis. First, all individuals have been risk-assessed based on the Domestic Abuse, Stalking and Harassment (DASH) Risk Identification and Assessment and Management Model (DASH, 2009 ). DASH was implemented across all police services in the UK from March 2009, having been accredited by the ACPO Council, now known as theNational Police Chief Council (NPCC). All police services and a large number of partner agencies across the UK use this common checklist for identifying, assessing and managing risk. The DASH index classifies individuals into one of three categories: standard risk, medium risk and high risk.
To measure the severity of crimes, we employ first the Cambridge Crime Harm Index (CHI, see Sherman, 2020 (link)). According to the CHI, the harm score for a crime is the default prison sentence that an offender would receive for committing it, if the crime was committed by a single offender with no prior convictions. For minor crimes that would instead result in a fine, the harm score is the number of days it would take someone with a minimum wage job to earn the money to pay the fine.
The second crime severity index is the Crime Severity Score index (CSS) taken from the ONS (Office for National Statistics, 2022b ). The CSS is a weighted index that reflects the relative harm of an offence to society and the likely demands on the police.
The collected police data are not the outcome of a randomized control trial, which means we need to exercise ex post statistical control for confounding variables that affect both the treatment group formation (i.e., probability of selection into the treatment group) and the treatment group outcomes. To illustrate this point, suppose such a variable is age; older people may be more likely to participate in the CARA treatment, and CARA recipients may show reduced recidivism only because older individuals have relatively lower re-offending rates rather than from any effect of CARA.
We use propensity score matching, a commonly used statistical technique (see, Rosenbaum and Rubin, 1983 (link)), to remove the effects that arise from confounding variables. This method is a quasi-experimental method that seeks to mimic randomization to overcome issues of selection bias that plague non-experimental settings. This method will allow us to provide a valid estimate of the intervention effect. The economic impact of the change in outcomes between the treatment and control group is based on this estimated effect and the calculations appear in section “3. Results.”
Operationally, PSM involves a three-step process.
Propensity score matching (PSM) matches individuals in the control and treatment groups. Because these individuals are the same (or almost identical) according to a set of observed characteristics, any differences in re-offending are attributed to the treatment. The matching is done by comparing the probability that an individual is assigned to the treatment group. Therefore, a matched pair comprises two individuals with the same probability of being assigned to the treatment group, yet one of them belongs to the control group. This probability is called the propensity score. The first step of PSM is to use logistic regression to calculate the propensity scores of each individual in the sample, the second step is to match these individuals according to their propensity scores. As the number of individuals in the control group is different from those in the treatment group, we use kernel-based matching. We selected the Epanechnikov kernel with a bandwidth of 0.06, which are common choices in the literature, see, e.g., Heckman et al. (1997) (link). PSM is standard in this literature, see e.g., Cox and Rivolta (2021) (link).