Risk

To effectively undertake risk-related activities (such as risk assessment and management), we need to be sure that we agree on the meaning of the term ‘risk’. Formal definitions of risk in the context of mental health service provision refer to a future harmful event and generally incorporate its severity, imminence and likelihood or probability. However, we cannot take for granted that the notion of ‘risk’ is applied consistently in clinical practice. A phenomenological analysis of the Royal College of Psychiatrist's report on the assessment and management of risk to others (Royal College of Psychiatrists 2016) found that ‘risk’ was used in a way that conveys variable meanings (e.g. harm, the possibility of harm, characteristics of possible harm, a state with changeable potential for harm, and different states of potential harm) (Nathan Reference Nathan, Gabbay and Boyle2021b). Given the potential for protean interpretations of ‘risk’, we believe it is essential that clinicians, academics and authors of practice guidance are clear about their use of this term. In this article, the term risk is used to describe the ‘possibility of harm’. This conveys that something may occur without necessarily suggesting that it is likely to occur (i.e. it is ‘possible’ but not necessarily ‘probable’) and it leaves room for the further characterisation of the possibility.

To be able to characterise future harmful events, it is first necessary to commit to the types of event under consideration. Guidance does not usually describe this step, but it would seem reasonable to at least consider types of harm suggested by expressed intentions (e.g. Tom's statement about throwing himself off a bridge), a recurrence of harm that has occurred before (e.g. an overdose), other forms of the same category of harm (i.e. other suicidal behaviours) and other categories of harm given a particular mental state (e.g. harm to others linked to Tom's angry hostility). Risk assessment convention dictates that the possibility of any identified plausible future harmful events occurring should be delineated. Therefore, clinicians need to understand different ways of describing the delineation (Box 1) and ensure that their use of a chosen descriptor is valid.

Prediction

A judgement about whether an event may occur in the future could be presented as a dichotomous prediction, i.e. that harm will or will not occur. Alternatively, judgements about the future could be categorised (e.g. low, medium or high risk). However, such categorical approaches have limited utility in informing clinical decisions so as to meaningfully reduce infrequent high-harm events (Mulder Reference Mulder, Newton-Howes and Coid2016).

Classic probability

The chance of a possible event occurring can also be characterised in terms of probability, which is used to describe the relative frequency of an event. This classic interpretation of probability is the fraction of the total number of possibilities of equal likelihood in which the event occurs (e.g. a 1/6 probability of tossing a three with a fair die). Since the assessment of possible future harm in clinical scenarios such as the case vignette described above does not involve a fixed number of outcomes of equal likelihood, classic probability is not applicable.

Frequentist probability

A related interpretation of probability is the frequentist view, which represents probability as a ‘long-run frequency over a large number of repetitions of an experiment’ (Blitzstein Reference Blitzstein and Hwang2019). This would produce the same probability of tossing a three with a fair die, but it could also be used to reach an informed probability for an outcome when rolling an unfair die (i.e. where the possibilities are not equally likely). Clinical scenarios such as the one presented in the vignette are unique instances that cannot be replayed. Thus, even putting aside the obvious ethical objections, allowing a series of instances to play would not be the same as rolling the same die in an unchanging environment. Therefore, frequentist probability is meaningless in a unique clinical situation. However, there remains a question about the applicability to practice of data-sets from comparable groups of patients. For instance, in relation to a discrete outcome such as suicide, one could use the positive predictive value of a suicide risk assessment model derived from group-level data as the probability of a patient identified as ‘high risk’ subsequently dying by suicide. However, once it is recognised that a patient presentation at a point in time is a unique instance and that the patient's future will be influenced by a unique set of countless environmental and psychological factors in constant interplay, the limitations of using a data-set to inform decision-making in a single clinical encounter become apparent (Nathan Reference Nathan, Gabbay and Boyle2021b).

Uncertainty

Economists have long recognised that probability in the classic or frequentist sense cannot be used for decision-making in the context of a single unique instance (Sakai Reference Sakai2016). Frank Knight reserved the term ‘risk’ for those scenarios in which the probability distribution of the outcome is known, and instead used the term ‘uncertainty’ for those scenarios in which the situation is so unique that probability distributions are not available (Sakai Reference Sakai2016). Critically in the context of mental health risk assessments, Knightian uncertainty seems more applicable since we cannot reliably distinguish what will occur from what will not occur. This uncertainty has two underlying components: epistemic uncertainty, resulting from a lack of knowledge, and aleatory uncertainty, resulting from random or chance events (Large Reference Large, Ryan and Carter2017). If it is accepted that one cannot accurately determine much of what is going to happen in the future (including events that might have a bearing on the occurrence of harm-related outcomes), then it follows that much of the uncertainty about a particular future harm-related outcome is aleatory (Large Reference Large, Ryan and Carter2017). The vignette illustrates that in such clinical settings the uncertainty arises in a scenario involving agents making interdependent choices with strategic dimensions. This raises as yet unanswered questions about whether game theory modelling might offer greater insights into likely outcomes. To do so, account would need to be taken of the role of emotions in decision-making.

Subjective probability

Another interpretation of probability that does not rest on an imagined or actual series of trials is subjective probability. In this case, probability is conceptualised as a feature of the person undertaking the assessment rather than of the real world and is quantified on a scale of the degree of belief (Biedermann Reference Biedermann, Bozza and Taroni2017). For instance, in the case vignette above, Amy felt that Tom should have a brief crisis admission, whereas the home treatment team practitioner felt otherwise. This discrepancy in opinion could be understood as arising owing to differences in both Amy's and the home treatment team practitioner's prior assumptions. Amy may have recently seen a very similar patient who subsequently died by suicide, therefore leading her to believe that Tom is presenting as higher risk warranting admission.

Propensity

One may also consider the properties of the entity that are causally responsible for the long-run frequency (such as the physical property of a rolled die). This is known as propensity. The propensity notion has advantages over the frequentist perspective because it can be used for a single event. Its use in mental health scenarios would also encourage a focus on the mental processes (the properties that are causally responsible for action) and, by extension, on consideration of factors that might influence these processes. In this way, talk of propensity shifts the focus from abstract disembodied concepts (such as probability) to the mind of the person.

Information gathering

Risk assessment should not be seen as a stand-alone clinical activity. It is heavily dependent on a comprehensive psychiatric assessment, which involves integrating information from an interview with, and examination of, the patient together with data from other sources (e.g. personal and professional informants, and the available records). Certain parts of that assessment are especially pertinent to judgements about the possibility of harm. These include psychological, behavioural or environmental occurrences, which have been found in group-level quantitative studies to be associated with an increased likelihood of future harm to self or others (American Psychiatry Association 2016; Franklin Reference Franklin, Ribeiro and Fox2017). For instance, these studies have demonstrated that some of the elements contained in the case vignette are associated with an increased likelihood of suicide: these include a history of overdose, comorbid psychiatric disorder, alcohol misuse, relationship breakdown and feelings of hopelessness.

Special attention should be given to the characteristics of violent or suicidal ideation (e.g. frequency and intensity) and associated behaviours (e.g. intent, planning, actions) (Borges Reference Borges, Nazem and Matarazzo2019). Suicidal ideation refers to thoughts about ending one's life, which may be categorised as ‘active’ (i.e. thoughts, and a plan, to die) or ‘passive’ (i.e. thoughts to die, but no plan) (Turecki Reference Turecki, Brent and Gunnell2019). Although the clinician should pay attention to any statements suggesting a pattern of thinking, feeling and planning that is relevant to the potential for future harmful actions, it should also be recognised that internal experiences relevant to future actions are not experienced as circumscribed and unchanging entities in the way descriptors such as active or passive ideation suggest. Similarly, taking a phenomenological approach to explore mind-based correlates of disturbed behaviour is likely to lead to a more fine-grained and case-specific causal understanding of the behaviour than is possible when working just at the level of diagnosis or symptoms (Nathan Reference Nathan, Whyler and Wilson2020).

If an assessment entailed merely compiling a list of case-based factors that theoretically may increase the likelihood of a high-harm outcome, it would be of limited use for decision-making in an individual scenario (American Psychiatric Association 2016). Instead, the relevance of these factors should be understood in the context of the person's current experiences and triangulated with other sources of information, such as a collateral history from a family member. Moreover, the assessment should attempt to uncover underlying psychosocial processes that explain how the factors interact to influence this person's behaviour (as explained below).

Interpretation

Despite the identification of factors that predict suicide or serious violence, it has not been possible to translate this literature into clinically useful predictive tools (Mulder Reference Mulder, Newton-Howes and Coid2016). Although a meta-analysis of prospective studies predicting suicide following self-harm found four risk factors with robust effect sizes (previous self-harm, suicidal intent, physical health problems and male gender) they were ‘unlikely to be of much practical use because they are comparatively common in clinical populations’ (Mulder Reference Mulder, Newton-Howes and Coid2016). The study's authors, who also reviewed a range of risk scales, concluded that ‘the idea of risk assessment as prediction is a fallacy and should be recognised as such’. Another meta-analysis concluded that no unique set of risk factors could be reliably linked to an outcome of suicide, with predictive ability found to have not improved over the past 50 years (Franklin Reference Franklin, Ribeiro and Fox2017). There are a number of reasons. First, risk factors are non-specific and are shared by many individuals; for instance, the presence of mental illness is considered a risk factor for a suicide, despite the fact that most people with mental illness will not die by suicide (Franklin Reference Franklin, Ribeiro and Fox2017). Second, suicide risk is influenced by a complex combination of interacting factors, with each individual factor only having a weak association with the rare outcome of suicide (Zortea Reference Zortea, Cleare and Melson2020). Third, although clinicians are tasked with determining risk over periods of hours or days, most studies are not focused on acute or short-term prediction and therefore, unsurprisingly, longer follow-ups within studies did not improve prediction power and, in some cases, significantly weakened it (Ribeiro Reference Ribeiro, Franklin and Fox2015). It should be mentioned, though, that a case has been made for the use of probability scores from a risk assessment tool to complement an individual needs-based assessment, although how the clinician uses the scores requires further empirical examination.

Believing that it is possible to categorise patients accurately may even be harmful. Potential problems include unnecessarily applying more restrictive approaches to all those categorised as ‘high risk’ and directing resources away from the ‘low-risk’ groups despite these groups containing most of those who go on to die by suicide (because they are much larger in absolute terms) (Mulder Reference Mulder, Newton-Howes and Coid2016). Additional problems may arise as a consequence of an innate tendency of humans to automatically assign people to broad categories and then to make predictions about them on the basis of the assigned category rather than person's unique characteristics (Liberman Reference Liberman, Woodward and Kinzler2017). In the risk assessment/management context, this may lead to the clinician inflexibly adhering to beliefs about the patient's assigned risk categorisation even in the light of emerging evidence to the contrary. As will be discussed below, such inflexibility can encourage a countertherapeutic stance. In forensic mental health services, risk categorisation is still used, but its effectiveness in reducing violence remains to be determined (Challinor Reference Challinor, Ogundalu and McIntyre2021).

Uncertainty versus predictions and probabilities

Although there may be an emerging consensus that categorical prediction has no place in an individual clinical risk assessment in general psychiatric settings (e.g. Graney Reference Graney, Hunt and Quinlivan2020), there remains a question about the role of probabilistic approaches. In other fields, such as economics and geopolitics, it has been demonstrated that the practice of actively thinking probabilistically is associated with more accurate subjective probability estimates of future events (Tetlock Reference Tetlock, Mellers and Scoblic2017). However, there is a difference between improving the accuracy of these estimates and the use of the estimates to inform clinical decisions. To use a probability estimate to inform a categorical decision (e.g. to admit the patient or not) requires a threshold to be applied, which inevitably turns the estimate into a category (and therefore subject to the problems of using predictive categories).

The concept of ‘uncertainty’, which acknowledges that the probabilities of the outcomes are unknown or meaningless (Park Reference Park and Shapira2017), more accurately represents the realm in which mental health clinicians undertake risk assessments and it encourages a focus on reducing uncertainty rather than predicting outcomes or quantifying the probability of their occurrence. Accepting that we are dealing with Knightian uncertainty, in that the probability distribution of future harmful events in the case of a given patient is unknown, reduces our need to think of the future in terms of predictions and probabilities. The focus should then shift from trying to compute the likelihood of outcomes to instead concentrating on characterising future scenarios of interest and the circumstances that would appreciably increase the likelihood of harm. In the psychiatric context, this can be facilitated by formulation, which brings together relevant factors in a way that produces an explanatory framework for the behaviour of interest. This is commonly done by assigning factors different explanatory functions, such as predisposing, precipitating, perpetuating and protective. Using formulation as an explanatory framework and scenarios to speculate about the future are important steps in approaches to risk assessment and management in forensic settings. Formulations can be further enhanced by exploring the proximal mental processes that are hypothesised to cause the behaviour (i.e. the propensity) (Nathan Reference Nathan, Whyler and Wilson2020). Existing explanatory models of suicidal and violent behaviour can serve as evidence-based frameworks to assist the clinician (Klonsky Reference Klonsky2018) (Box 2 and Table 1). Such models describe how combinations of many risk factors create pathways towards the emergence of harmful ideation and behaviour, and reserve particular focus, as a result of being set within the ideation-to-action framework, on the factors that govern the transition from ideation to action (Zortea Reference Zortea, Cleare and Melson2020).

BOX 2 Explanatory frameworks of proximal mental processes related to suicidal behaviour

Interpersonal theory of suicide (IPTS) (van Orden Reference van Orden, Witte and Cukrowicz2010) – Thwarted belongingness (unmet need to belong) and perceived burdensomeness (perception of being a burden to others) lead to suicidal desire. Exposure to painful and provocative events (e.g. childhood adversity, adult trauma, self-injury) causes habituation to fear and pain, which can result in the capability to enact suicide.

Integrated motivational-volitional model (IMV) (O'Connor Reference O'Connor and Kirtley2018) – A tripartite model in which (1) the pre-motivational phase, comprising life circumstances and events, provides a contextual background for (2) the motivational phase, when defeat and humiliation appraisals lead to a state of entrapment in which suicidal behaviour is seen as a salient solution to life circumstances and suicidal ideation and intent develop, followed by (3) the volitional phase, in which the suicidal behaviour occurs. Progression through these phases is influenced by moderators (e.g. social problem-solving, memory biases and rumination, thwarted belongingness, burdensomeness, social support).

Three-step theory (3ST) (Klonsky Reference Klonsky2018) – Step 1: the combination of pain that is experienced as punishing (particularly psychological pain consequent on aversive stimuli, such as conflict, loss, physical ill health) and hopelessness (about the pain diminishing) lead to suicidal ideation; step 2: suicidal ideation is heightened by disrupted connectedness (primarily to other people, but also to a job, project, interest, etc.); and step 3: progression from ideation to action depends on suicide capacity, which is made up of the dispositional (e.g. pain sensitivity), acquired (e.g. habituation to experiences associated with pain) and practical (e.g. knowledge of and access to lethal means).

Fluid vulnerability theory (FVT) (Rudd Reference Rudd and Ellis2006) – Takes account of the variable nature of suicide risk over time (such as rapid changes in ideation) by delineating chronic and acute suicide risk. Acute episodes of risk are time-limited and driven largely by situation and contextual variables (e.g. severity and mix of current symptoms, life stressors, substance misuse, access to method). Chronic (or baseline/residual) risk describes an enduring individual vulnerability for suicidality or an acute risk episode and is based on a disparate range of factors, including adversity, genetic history, biology/physiology, and cognitive, affective and motivational dispositions.

TABLE 1 Explanatory framework of mental processes related to violent behaviour

Decision-making

Although any categorical or dimensional outputs of risk assessment methods may not be usable in individual instances of clinical decision-making, the constituent elements of the risk assessment may still have some utility for risk management. Risk management involves addressing current modifiable factors that the explanatory formulation suggests increase the likelihood of high-harm behaviour, including any specific treatment needs (e.g. depression or substance misuse) (Large Reference Large, Ryan and Carter2017). Ways of bolstering factors that the formulation suggests reduce the likelihood of future harm should be explored with the patient. The management plan should also consider early warning signs for a high-risk mental state and steps to be taken in this event. Specific safety or crisis plans (which focus on stressors, warning signs, self-management strategies, social support and crisis resources) have been shown to reduce subsequent suicidal behaviours (e.g. Nuij Reference Nuij, van Ballegooijen and de Beurs2021).

Although it has to be accepted that the current empirical evidence undermines the case for the routine clinical use of categorical risk predictions, it does leave an outstanding question about how clinicians should take account of risk when faced with options that can only be implemented in categorically distinct forms. In the vignette presented above, one of the issues under consideration was admission to hospital. Clinical experience would suggest that the decision is often framed as being dependent on whether or not the patient is going to imminently act on urges to cause harm (to self in the vignette) but, as has been explained, this cannot be accurately predicted. It is possible to conceive the use of a probability estimate, but it would require the data-set from which the estimate is drawn to be relevant to the scenario (i.e. from a sample of patients assessed following an overdose rather than an in-patient sample about to be discharged). Also, to use a number from a continuous scale to make a categorical decision requires the establishment of thresholds (with or without the scope for justified discretion). Before introducing such an approach, further empirical testing would be necessary.

An alternative approach would be to use the unique case-based explanatory formulation in collaboration with the patient to consider the factors that increase the likelihood of the behaviour and potential ways to address modifiable factors so as to reduce that likelihood. It would still be necessary to come to a categorical decision, but this approach has the advantage of focusing attention on what can be done to adjust the modifiable circumstances (and this may include the use of therapeutic skills during the assessment), which in turn may assist the clinician in combatting their own automatic and biased categorical assumptions.

It needs also to be recognised that in practice the system places demands on the clinician that are sometimes in competition. Although the imperative to help the patient's emotional well-being may not conflict with the requirement to reduce the likelihood of high-harm outcomes, sometimes it does. For example, although admission may be thought necessary to manage dangerous levels of self-harm, in some cases the in-patient environment may exacerbate some of the underlying processes contributing to self-harm. For example, the use of medication and/or physical restrictions to address acute mental state changes may reinforce certain pre-existing processes contributing to a tendency to self-harm (such as difficulties understanding and reflecting on the mental states of others, which may lead to a dependence on the overt physical representations of the intentions of others provided by medication and physical restrictions). Therefore, the clinician needs to consider the possibility that the way of managing the potential for harm in the short term (i.e. admission) will further compromise the patient's well-being. Also, there may be service-level pressures (which can be manifest in interpersonal tension between ‘referrers’ and ‘gatekeepers’) to avoid admission, which are often related to the need to manage a finite resource under pressure. Thus, the clinician's final decision is influenced by judgements about future unpredictable harms and benefits within a complex system of pressures and relationships.

Interaction

If risk assessment is framed solely as a process of gathering and interpreting information about the patient to reach a judgement about the possibility of future adverse incidents, then the discussion with the patient is liable to become transactional, in that the primary goal is to elicit information. To the contrary, the interaction should be seen as an opportunity to reach a shared understanding of risk and to positively influence the patient's mental state and their representations of mental healthcare professionals in a way that may have a bearing on their future behaviour (including behaviour that has the potential to harm themselves or others). The clinician should adopt a compassionate and collaborative stance in which they actively and supportively listen with the aim of enhancing understanding (or reducing uncertainty) rather than reaching a categorical position about risk (Large Reference Large, Ryan and Carter2017).

There may be a legitimate difference of view between the clinician and patient about the next steps (e.g. whether to admit a patient who is feeling suicidal), but the clinician needs to be careful that their explanation for their decision does not constitute invalidation of the patient's experiences. The false belief that the patient can be definitely allocated to a risk category may lead the clinician to have an undue sense of confidence in their judgement, and then convey this false confidence to the patient in a way that invalidates the patient's experiences and feelings. There is empirical evidence to show that a collaborative assessment that emphasises respect, compassion, curiosity, a desire to help, flexibility and the relational nature of the assessment is associated with symptom reduction and patient self-enhancement (e.g. Durosini Reference Durosini and Aschieri2021) and it has been proposed that a positive therapeutic relationship may reduce risk (Royal College of Psychiatrists 2016).

Learning and investigating

The approach to learning from untoward incidents is relevant to the practice of risk assessment because in making decisions in acute psychiatric scenarios, clinicians are influenced not only by the clinical information available to them, but also by the wider context in which they practice (Nathan Reference Nathan and Wilson2021a). Such learning, typically through investigations, takes a deterministic perspective that does not readily apply within complex systems (such as mental health services), in which the underlying determinants of change are uncertain and non-stationary.

Although scrutinising the past in light of the outcome seems the best way to reach an explanation of why the outcome occurred, it can lead to problems when judging whether the decision was the correct one or not. The usual approach to investigating an incident (e.g. by an internal investigation, an external review or the coronial process) is to work back from the outcome to identify factors and decision points that appear to have a causal link with the outcome. This approach encourages decision-making to be tested in light of the outcome, i.e. would the outcome have been different if a different decision had been made? It is well recognised that looking back from the outcome makes the assessor susceptible to hindsight bias, which leads to an overestimation of the foreseeability of the outcome (Roese Reference Roese and Vohs2012) and that the resulting investigations can contribute to a culture in which clinicians feel that they are liable to be blamed for unpredictable outcomes (Munro Reference Munro2019). However, more fundamentally, the decision maker is in effect being tested against a standard that is almost impossible to meet. Not only do they not know the outcome, but as articulated above the empirical evidence also suggests that even if they had fully considered the relevant risk factors, that outcome would have been nigh on impossible to predict in an individual case. Another way of illustrating the contradiction is to consider the decision not only in light of actual events but also in the counterfactual scenario in which the outcome did not occur. Concluding that it was reasonable in the latter scenario but not in the former means that the clinician is placed in the unenviable position of being assessed by a standard that is not known at the time of making the decision and only becomes known after a future unlikely and unpredictable outcome has occurred (Bhandari Reference Bhandari, Thomassen and Nathan2022).

What is the relevance of this to practice? At the time of making a critical clinical decision, as well as thinking about the possibility of future harm enacted by the patient, clinicians’ decisions are also influenced by an in-the-moment contemplation of an unfavourable future appraisal of their decision-making in the event of a serious adverse event occurring (even though the event is unlikely) (Nathan Reference Nathan and Wilson2021a). The issues can be illustrated by a development of the scenario presented in the vignette above. Although Amy remained of the view that admission was indicated, there had been previous tension between her and the home treatment practitioner and she did not have the confidence to contest the decision to decline admission. She also had on her mind thoughts about a member of her team who was still off work for stress due to the experience of the way she was questioned in both an internal investigation and coroner's inquest in a separate case 6 months before. When Amy told Tom that he had to go home, he abruptly left the hospital. In Amy's clinical entry she justified her decision not to pursue the option of admission by stating that although Tom ‘threatened’ to throw himself off a motorway bridge this appeared to be for ‘secondary gain’ (to secure admission) and she thought that because he was talking about events in the future the risk of suicide was not high. Thus, although in reality clinicians are unable to reliably predict the outcome, they are prone to hold in mind an approach that relies on a deterministic predictive model (in which an outcome is an inevitable result of the antecedent causes), further encouraging the spurious categorisations of patients as high or low risk, the use of such categorisation to allocate resources and the adoption of defensive practices (Munro Reference Munro2019). Amy would, she thought, have felt more able to take a different approach to the dilemma she faced if it was clear to her that in the event of a future untoward incident, the investigation would not use the outcome as a basis for judging her actions and account was taken of the real-time complexity of the situation. She would have felt less motivated to use a spurious risk categorisation to defend herself against what she imagined, from the perspective of a hindsight-armed investigator or coroner, could look like a ‘missed opportunity’. More importantly, by feeling able to acknowledge with Tom (and in her clinical entry) that she remained concerned about the risk, the encounter would have been more validating and there would have been greater opportunity for the collaborative exploration of alternative options to admission, which as noted above may have reduced risk.

Therefore, the issue is not just agreeing how to conceptualise risk assessment, but also agreeing how to conceptualise the system in which the risk assessment activity occurs. Rather than conceptualising a deterministic system (i.e. one in which the future state of every aspect of the system can be described), risk assessment should be characterised as an informed process to reduce uncertainty within a complex system of unknowable future states. More research is needed before firm recommendations can be given about alternative models to learn from adverse incidents, but possible approaches include appraising any decision of interest in light of a counterfactual scenario in which the outcome did not occur, appraising decisions against pre-agreed standards, appraising decisions without knowledge of the outcome and/or focusing on potential latent factors (i.e. systems, processes and training) rather than just linear cause-and-effect ones (Bhandari Reference Bhandari, Thomassen and Nathan2022).