Escape and absconding among offenders with schizophrenia spectrum disorder – an explorative analysis of characteristics

In the mental health system, escape is defined as gaining freedom by breaking through the secured perimeter, in forensic psychiatric hospitals often the outer wall or fence of a ward [1, 2], while absconding is defined as gaining freedom by evading the supervision of staff during a controlled opening outside the ward or hospital [1]. As forensic psychiatry is defined as “a subspecialty of psychiatry in which scientific and clinical expertise is applied in legal contexts involving civil, criminal, correctional, regulatory, or legislative matters, and in specialized clinical consultations in areas such as risk assessment or employment” this is a crucial area of research and daily clinical practice in this field.

Both escape and absconding from closed psychiatric wards, and especially from secure forensic hospitals, can be serious events leading to potentially grave outcomes [1‐5]. These can directly affect the patient concerned, such as overdosing [6‐8], self-inflicted injury [8‐10], attempted suicide [10‐14] or completed suicide [15, 16], prolonged rehabilitation and treatment [17], medication non-compliance [17], violence against others [4, 9, 14] or (re-) offending [8, 13, 15, 18‐22].

The reputation of institution and staff can also be damaged by such incidents and subsequent negative media coverage regarding mental health and security may lead to increased stigma, reduced confidence in the particular institution or the health care system as a whole, anger and guilt among staff, stress and anxiety among patients [4, 23‐25].

Despite those detrimental effects, research on this topic is rather sparse [1, 2]. Recent reviews have identified 39 studies on escape and absconding in general psychiatry [2, 17], but only nine on the subject in forensic psychiatry, although the consequences are particularly severe in the latter setting [23].

Most of the relevant research on escape and absconding was conducted by Bowers [9, 26‐28], who focused primarily on the analysis of the characteristics of affected patients. Subsequent studies on this topic were based on these characteristics and were predominantly undertaken in the United Kingdom and Australia. The diagnosis of schizophrenia spectrum disorder (SSD) has been identified as one of the most influential characteristics for escape or absconding with incidences of up to 71% in patients examined [1, 4, 5, 17, 23, 24, 27‐33]. Further risk factors for escape or absconding included: younger age [1, 3‐5, 9, 29, 33‐36], male sex [1, 4, 5, 29, 30, 33, 36‐38], being unemployed [4, 37, 39], being unmarried [1, 5, 7, 10, 36, 39], homelessness [38], the number of diagnoses [38], having a personality disorder [1, 4, 5, 11] or affective disorder [4, 5, 9, 33, 36], prior convictions [4, 17, 19], forensic psychiatric treatment in the past [1], absconding in the past [1, 5, 23, 24, 30, 32, 35, 39], compulsory detainment [1, 5, 38], treatment non-compliance [1] and medication non-compliance in the past [5], exacerbation of symptoms [19, 28, 31], aggressive behavior [24, 38], alcohol or drug abuse [1, 4, 5, 23, 24, 33, 38‐40], a history of sexual abuse [1], not having a history of self-inflicted harm [38], suicidal behavior [23], shorter duration of current hospitalization [3, 4, 33, 34, 38, 41] and also longer duration of current hospitalization [23, 30].

Although past research has provided important insights into the topic, certain weaknesses are apparent: (1) Previous studies included heterogeneous patient populations with different diagnoses. Although individuals diagnosed with SSD constitute the majority of escapees and absconders, there is no study to date that has examined these patients exclusively (2). The investigated features are the result of the valuable work of Bowers and his colleagues and were reproduced in subsequent studies. However, new and potentially more specific factors could not be uncovered so far. (3) The statistical analyses of earlier studies were mostly based on null hypothesis significance testing (NHST), in rare cases on linear regression methods, which was already recognized as a problem in 2000 [42], yet received only limited attention in further studies. The problem with examining multiple single variables with NHST, known as multiple testing [43, 44], is that each statistical test is associated with a false positive rate (type I error) and continues to accumulate with each additional test. This results in variables being falsely identified as significant and is further exacerbated when there is no correction for multiple testing [45, 46]. Another limitation of the NHST is the insufficient consideration of interactions between the variables and that no evaluations of the quality of statistical models are provided. Simply looking at the significant variables may produce models and interpretations of limited informative value. Therefore, the aim of this study is to investigate the phenomenon of escape and absconding employing a complex data set of 370 offenders with SSD and over 500 potential influential variables via machine learning.

An overview of the performance parameters of the different calculated models during the nested resampling procedure can be found in Table 1. With a balanced accuracy of 76.7%, (95% CI = [68.4, 82.7]) and an AUC of 0.88 (95% CI = [0.81, 0.95]) naïve bayes outperformed all other ML algorithms.

The quality of the final model in the validation step is shown in Table 2. As expected, the balanced accuracy of 71.1% (95% CI = [58.5, 83.1]) and an AUC of 0.75 (95% CI = [0.63, 0.87]) were less than the results of the initial training model, but still meaningful. With a sensitivity of 88.2% (95% CI = [87.9, 88.5]) over three-quarters of patients who did not escape or abscond were identified correctly, with a specificity of 60% (95% CI = [59.2, 60.8]), less than two-thirds of patients, who escaped or absconded were detected correctly.

The absolute and relative distribution of the 10 most influential variables among the whole dataset can be seen in Table 3. In addition to age at diagnosis, age at admission and number of index offences, the other influential variables covered circumstances related to the current hospitalization. Testing for multicollinearity showed no dependencies between the variables (detailed results see supplementary materials).

A one-sided tornado graph comparing the relative importance of the identified variables during model validation is presented in Fig. 2. It shows the effect on the output variable (patient escaped/ absconded and patient did not escape/ abscond) by varying each predictor variable at a time, keeping all the other predictor variables at their initial values. The x-axis represents the relative variable importance, the y-axis each variable - the wider the bar, the more impact the variable has on the model and the outcome. Consequently, the predictor variables are ranked from the most influential to the least influential.

The aim of the study was to identify the most important characteristics of offender patients with SSD who escaped or absconded from a forensic hospital by means of ML and to describe the best possible statistical model. The identified variables were age at diagnosis of the SSD and age at admission, time spent in current hospitalization, problematic behavior during current hospitalization (e.g., forbidden intake of drugs/ alcohol, breaking of rules, dissocial behavior), amount of index offences, PANSS score at admission and discharge (note: if the patient did not return after the event of escape or absconding, the last PANSS before the event was taken) and dosing of neuroleptic medication (daily cumulative olanzapine equivalent at discharge). The final model utilizing these variables achieved a balanced accuracy of 71% at an AUC of 0.75 with a sensitivity of 80% and a specificity of 60%.

The escaped or absconded offender patients suffering from SSD shared several characteristics with escapees or absconders identified in previous studies in mixed populations. Our results confirm studies showing that a younger age plays a central role in escape or absconding [1, 3‐5, 9, 27, 29, 33‐36]. Not only the age at the time of the incident was important as described in previous studies, but also a younger age at the time of the initial diagnosis of SSD. A shorter hospitalization period could also be confirmed [3, 4, 31, 33, 34, 38] indicating that escape and absconding arise in an earlier phase of treatment. Although, in contrast to previous investigations, no increased incidence of diagnoses of substance abuse was found [1, 4, 5, 23, 24, 33, 38‐40], there was a higher incidence of substance abuse during the current forensic hospitalization. Interestingly, this behavior was embedded in a general behavior of dissociality and a proneness to rule breaking. In this context, the newly identified variable “number of index offences” may be a possible indicator for an increased propensity to breach rules even before treatment. However, a difference in the frequency of diagnosis of dissocial personality disorder between the groups was not apparent, since diagnosing a personality disorder in the presence of SSD is excluded according to ICD coding.

Although the evidence on the relationship between escape or absconding and symptoms of SSD is scarce [25], previous studies have reported that symptoms were exacerbated at the time of escape or absconding [19, 28, 31]. Based on the objective scores of the PANSS over the course of treatment, our findings indicate that patients who escaped or absconded not only had an acute exacerbation of symptoms at that time, but they generally displayed a more severe course of the disorder. Presumably, this is also reflected in a higher demand for neuroleptic drugs, which was found to be an influential factor in our sample. Our model did not identify any variables regarding childhood or adolescence and, with the exception of the two variables mentioned above, also no parameters from psychiatric or criminal history [1, 4, 5, 17, 19, 23, 24, 38]. Hence, our analysis indicates that past patient information is less relevant than information on the current hospitalization.

In summary, we found that offender patients with SSD who escape or abscond tend to be younger, have a more severe level of pathology and are more likely to exhibit dissocial behavior.

In general research on populations of offenders with SSD, especially in the context of violence, age [68‐74], dissocial patterns [68‐75], substance abuse [75‐77] and psychotic symptoms [78‐85] have also been identified as risk factors and dominate the field of forensic research on SSD and offending. Escape and absconding are rule-violating behaviors and thus perhaps a facet of dissocial behavior patterns that are interwoven with psychotic symptoms in a complex multifactorial framework that has not yet been sufficiently understood. It is possible that the group of offenders suffering from SSD comprises a subgroup of individuals who are particularly affected by these factors. For example, the framework by Hodgins’ [69] postulates a subgroup (“early starters”), which is characterized by young age, dissocial and rule breaking behavior and drug abuse. They may also be more prone to problematic behavior during hospitalization and more likely to escape or abscond, which should be investigated in future studies.

Although the goal of the study was not to develop a tool to identify patients who flee or escape, the usefulness of the current model and subsequent opportunities for the potential development of such a tool need to be discussed. This will be illustrated by a cost matrix applied to the results of the validation data set. Patients correctly identified as non-absconders/ non-escapees (TP - 60 patients) or absconders/ escapees (TN - 9 patients) have no direct negative impact on the hospital and therefore generate zero cost. Patients who are incorrectly identified as Absconder/Escapees (FN - 8 patients) may receive more extensive treatment and be unnecessarily restrained. This results in costs of 20 units per patient. Patients who are misclassified as non-absconders/ non-escapees and then nonetheless escape (FP - 6 patients) pose a massive risk to society and themselves [4, 23‐25]. Here we postulate a cost of 100 units per patient. Calculating the total cost for our current model results in: 60 × 0 + 9 × 0 + 8 × 20 + 6 × 100 = 760 units. A perfect model with 100% specificity and sensitivity (0 FN/ 0 FP) results in no cost: 68 × 0 + 15 × 0 + 0 × 20 + 0 × 100 = 0 units. If the model correctly identified a patient who was misidentified as an absconder/escapee (sensitivity increase), this would result in the trade off of misclassifying a patient as a non-absconder/ non-escapee, resulting in an overlooked escape (specificity decrease) and in a cost of: 61 × 0 + 8 × 0 + 7 × 20 + 7 × 100 = 840 units. If we now turn these considerations around and increase specificity with a loss of sensitivity, i.e., identify absconder/escapees with more efficiency, this results in a cost of: 59 × 0 + 10 × 0 + 9 × 20 + 5 × 100 = 680 units. In summary, a decrease in specificity leads to an increase in undetected absconders/ escapees and a massive cost. In other words, for one missed absconder/ escapee, five patients may be erroneously treated too restrictively. Thus, these considerations show that detection of actual absconders and escapees are essential for model building. It also demonstrates that unbalanced data can lead to costly consequences. But it is not only these statistical considerations that must be considered when developing a potential tool. How expensive is it to treat a patient too restrictively, for example, and how much more expensive is it to have a patient drop out and potentially commit new crimes? At what level can we talk about acceptable risk and what is the minimum level of sensitivity and specificity that must be achieved? Rather complex questions that affect different areas like health and security policy, ethics must be addressed when developing and using a tool for risk assessment.

The results of the present explorative study may, nevertheless, provide benefits for clinical practice. By identifying a potentially problematic subgroup early in the course of hospitalization, the needs and risks of said subgroup can be addressed more effectively. A more individualized treatment of symptoms and an expansion of therapy to address dissocial behavioral patterns could possibly reduce the rate of escape and absconding.

However, since most of the risk assessment tools employed today serve as the basis for a variety of decisions about effective punishment and treatment [86], care must be taken to ensure that patients are not stigmatized or receive inferior treatment due to certain parameters like race or gender. Courts use such tools to assess the likelihood of recidivism or escape of pretrial detainees or offenders in bail and probation proceedings, or to set bail amounts [87]. ML methods tend to shift the trend of risk assessment toward prediction rather than intervention, as prediction-oriented risk assessments do not take into account the criminogenic nature of the criminal justice system itself, although the socially useful purpose of risk assessments should be directed more toward their utility as diagnostic and treatment guidance tools [86]. Bias can enter the design of a statistical model as well as the interpretation since the interpreter may more or less unconsciously incorporate personal political orientations, values and opinions when interpreting results [87]. In line with other authors [86‐88], we emphasize the importance of providing decision makers, such as judges, with additional information about the limitations immanent in risk score predictions and encourage further efforts to better understand how risk assessment tools interact with biases and beliefs held by the individuals who rely on them.

In conclusion, although the performance parameters of our model seem to be reasonable, this sensitive topic with consequences for the medical treatment of humans is far from being an automated machine selection process. To achieve this, high sensitivity and specificity rates would be required. In this study, ML must be seen as an advanced statistical method for retrospective differentiation of individuals rather than a predictive modeling technique.

For the first time it was possible to create and describe a model using innovative ML analysis to identify influential factors related to escape and absconding in forensic patients with SSD. Patients who escaped or absconded were generally younger and showed more pronounced pathological manifestations. They exhibited an overall more problematic and rule-breaking behavior. Early identification of these patients could help to prevent problematic events from occurring in the first place by providing specifically tailored treatment strategies for these patients.