Are homicides by mentally disordered persons a major public health problem?

There are about 500 homicides per annum in England and Wales. Around 50 now receive a verdict of manslaughter under Section 2 of the Homicide Act 1957, that is, a plea of diminished responsibility is accepted (Reference Taylor and GunnTaylor & Gunn, 1999). Less than 50% of these probably suffer from a psychosis (Reference DellDell, 1984). These Home Office statistics are consistent with numbers in a recent report from the National Confidential Inquiry into Homicides and Suicides (Department of Health, 1999) which showed that less than 20 homicides were committed by persons with a psychosis per year, including 10 with schizophrenia who had ever had contact with mental health services. By comparison, in the UK in 1994 there were about 5000 suicides, 500 deaths from fire, 4000 deaths from accidental falls, 280 deaths from drowning, 4000 deaths from motor vehicle accidents and 12 000 deaths from other accidents and adverse events (World Health Organization, 1998). Why no inquiry for each of these? Especially frightening to the public is the prospect of being killed by a stranger with psychosis. In fact the risk of this is around the same as that of being killed by lightning — about 1 in 10 million.

Can these homicides be eliminated?

There is no evidence that homicides by people with mental illnesses have increased over the past decade; indeed Section 2 manslaughter homicides have fallen (Reference Taylor and GunnTaylor & Gunn, 1999; Reference Szmukler, Thornicroft and HollowaySzmukler et al, 1999). The alleged relationship with ‘community care’ is spurious. However, each homicide by a person with a mental illness is trumpeted as yet another example of the failure of ‘community care’. An assumption reigns, among the media and politicians at least, that all such homicides are preventable, despite the fact that every country has, and has always had them. For some reason, ours has become terrorised by them. I find it embarrassing when colleagues from other countries show surprise and ask why. They know that preventing homicides is like stopping accidents of wars. We do our best to try to reduce their likelihood, but no-one has ever stopped them. They are part of the human condition.

Are homicides by people with mental illnesses predictable?

Is it reasonable to criticise a psychiatric service for failing to prevent the unpredictable? If the annual prevalence of psychosis is about 4 per 1000 population as estimated by the Office of Population Censuses and Surveys (1995), it can be calculated that a homicide occurs once in about every 10 000 ‘psychosis-years’. If there were a predictive test for homicide by a person with psychosis with a widly unrealistic ‘sensitivity’ (the proportion of actual homicides predicted as homicides by the test) of 0.9 and a ‘specificity’ (the proportion of actual non-homicides predicted by the test as non-homicides) of 0.9, then the positive predictive value of the test would be negligible (for every homicide accurately predicted, there would be about 2000 false positives). The result would be of the same order for a test with a ‘sensitivity’ of 0.1 and a ‘specificity’ of 0.99. Risk factors for violence by mentally ill persons are common, but homicide is extremely rare. If all persons with risk factors were treated as potential perpetrators of homicides we would deprive many thousands of their liberty to (possibly) avoid one death. There would still be, of the total homicides committed by persons with a mental illness, a large proportion, indeed probably larger, from the ‘low risk’ patients since there are so many more of them than those who are ‘high risk’ (Reference Shergill and SzmuklerShergill & Szmukler, 1998).

The patient as an ‘automaton’

Inquiries, perhaps because of their terms of reference focusing exclusively on mental health services, adopt a model of responsibility that is grossly oversimplified and distorted by retrospective analysis. The offender patient is seen as lacking agency, behaving as an ‘automaton’, like an aeroplane out of control. The patient is no longer a person with feelings, hopes, intentions or with the capacity for choice. For example, it seems patients never exercise choice in making decisions about taking medication or not, or misusing dangerous substances. What about intercurrent and unpredictable events, the kind that could happen to all of us, that pushed the patient in the direction of violence? Was there provocation by others? These are ignored — the patient becomes a mere cipher. There is little or no attempt to explain the violence as ‘normally’ understandable in the circumstances.

Another aspect of the automaton model of the patient is a ‘cancelling out’ of pre-illness violence related to personality or social factors. In our inquiry, for example, a history of serious violence antedating the illness passes without comment. The assumption is that the final act was entirely due to mental illness; there is nothing left which need concern us. On the other hand, research shows that the strongest predictors of violence in the mentally ill are previous convictions, gender and a range of social factors, none of which are ‘treatable’ (Reference WesselyWessely, 1997).

Absence of a role for other actors

The focus on services again excludes the role of other actors, including possible risk-taking by victims, or a failure to act by others who were in a position to help. In our inquiry, the victim visited the patient at 3 am despite advice from friends not to do so. Earlier that night, he had threatened (with scissors) another young woman he had known. She called the police at about 10 pm, but they did not follow-up on the incident because she did not want to press charges (a curious paradox is here passed over unnoticed: the agency with a responsibility for preventing crime, but not for establishing a therapeutic relationship is held blameless, while the agency with a primary responsibility for treatment, but not for preventing crime is regarded as not having exercised sufficient control of the patient!)

Being responsible for the behaviour of another person

The surgeon knows within narrow limits how an artery will respond to being severed and thus rightly carries responsibility for the ‘behaviour’ of the artery during surgery. Can a mental health team carry the same responsibility for the behaviour of another person, an agent with a mind of their own? Take an analogy with the crew of an aeroplane. It is recognised that teamwork is a major factor in avoiding accidents. Would one hold them responsible in the same way if their plane had a mind of its own? What if the plane felt aggrieved at the way the pilot had landed the previous day, or felt humiliated by being given a distant parking bay at Heathrow. Would we still see the crew as solely responsible for the behaviour of the plane if it decided not to ascend as directed?

Responsibility for the behaviour of another person is, I believe, unprecedented in medicine. It occurs in social work where child protection teams are held to some extent responsible for the behaviour of an abusing parent. The argument presented above would make this unreasonable in the same way. However, the expectation of the mental health team is some way less rational still. The child protection team knows who the victim will be, and to some degree the range of behaviours, usually fairly narrow, of the potential victim and offender. The mentally ill person's potential victims are usually uncertain, and relationships complex.

Retrospective causal indeterminacy

Inquiries seek causes based on a detailed reconstruction of past events, in turn based on information from a variety of sources. Links between events in a retrospective narrative are usually indeterminate. One common approach is to examine a set of requirements placed on a service, such as adherence to the Care Programme Approach (CPA; Department of Health, 1990). Deficiencies in this regard become the subject of extensive criticism, but it is almost impossible to show that such deficiencies were related to the final outcome. For example, while CPA meetings of all involved members of a multi-disciplinary team, representatives from other agencies, together with the patient and carers may not have occurred regularly, would such meetings have led to different clinical decisions? For most decisions it is impossible to know. It is also of interest that what started out as ‘guidance’, such as the CPA or discharge planning, becomes a ‘standard’ or test, even though a recent Cochrane Review concluded that ‘case management’ methods like the CPA have a very limited effect on patient outcomes (Reference Marshall, Gray and LockwoodMarshall et al, 1997).

It is also difficult to show that any decision was significantly related to the final outcome, particularly if we reject the ‘automaton’ model of the patient. For example, a major conclusion of our inquiry was that the patient should have been recalled to hospital when he refused further depot medication (10 months before the homicide), even though he showed no symptoms of psychosis, agreed to oral medication and agreed to closer monitoring. If greater coercion had been used then, how can we know that the patient might not have become even less cooperative, more embittered and thus more dangerous?

Hindsight bias

Hindsight, despite the inquiry team's best efforts, is pervasive. With hindsight an outcome begins to look inevitable; a plausible chain of causes can easily be traced backwards through time. One loses an awareness of the multitude of possibilities presenting themselves at any moment in ‘real’, forward-moving time, and the immense range of possibilities consequent upon this range of choices. Also easily lost is an awareness of the extensive range of causes and sequences that might have led to the final outcome. However, one retrospective sequence is usually chosen, punctuated by apparently ‘critical’ points; the argument then becomes: ‘if x had done y, or not done z, this need not have happened’. This phenomenon of selectively focusing on data consistent with the previously revealed outcome has been called ‘creeping determinacy’ in the hindsight bias literature (Reference Hawkins and HastieHawkins & Hastie, 1990).

I identified at least two variants of hindsight, the ‘blatant’ and the ‘subtle’. The ‘blatant’ is illustrated by some major conclusions in our inquiry — the patient should not have been allowed to stop his depot medication; if he had lived in supervised accommodation he could not have been visited by the victim in the early hours of the morning; a home visit should have been made on the same evening following his presentation to out-patients, where he came seeking his consultant and not planned for the following morning (the homicide occurred that night). The ‘subtle’ is more difficult to discern. It determines how evidence that is ambiguous is weighed in the balance. Our inquiry concluded that the patient had shown signs of relapse for some months before the homicide. This was based on reports of odd behaviour by family members (which the inquiry conceded presented some difficulties) and information obtained by the inquiry, but not available to the treatment team at the time, from police and the publican where the patient used to drink. On the other hand the treatment team, despite more than a dozen contacts with the patient over the preceding two months, could find no clear evidence of relapse, although they were concerned about the possibility, and specifically looked for signs. Furthermore, the family of the victim who had contact with the patient for some three months before the homicide, apparently had no idea that the patient suffered from a mental illness. Indeed, on the day the report was released, the family complained to one newspaper that they should have been warned he was a “dangerous schizophrenic”; they had not suspected. With hindsight knowledge that the patient did kill, the inquiry concluded the patient had relapsed, but this was missed by the treatment team. This subtle influence of hindsight is insidious and pervasive.

Inquiries ignore the evidence relating to ‘hindsight bias’, mostly discussed in relation to court hearings in negligence cases (Reference Hawkins and HastieHawkins & Hastie, 1990). Knowledge of the final outcome in a particular case, increases the estimated likelihood in the juror's mind of such an outcome against other equally likely, or more likely, outcomes. Even being warned about such a bias fails to reduce it. More active measures are required, for example, keeping the ultimate outcome secret, or forcing jurors to find reasons from the facts of the case for alternative outcomes, but even these may be relatively ineffective (Reference Stallard and WorthingtonStallard & Worthington, 1998). Hindsight bias may also be stronger the more data there are available (Reference PenningtonPennington, 1981); a 500 page report provides plenty.

What can be learnt?

What can be learnt from a single case study, no matter how detailed, is very limited: even more so when the case is atypical in the extreme, as is a homicide. Given the problems in ascribing causal relationships discussed above, nothing definitive can be concluded abut why the homicide occurred. Furthermore, we cannot know how representative the practices observed are of the psychiatrist's or team's usual practice, or how the service's policies and procedures are generally implemented. To know how well a team or service functions, an audit against an established set of standards is required. A single case study can, at best, generate hypotheses for testing. Does adherence to the elements of the CPA result in fewer homicides by mentally ill patients? Carefully mounted, complex research studies are required to answer such questions. For a rare event like homicide, a case—control study is probably the most practical option if one wants to discover what treatment factors are associated with such an outcome. For the astonishing cost of our inquiry, over £750 000, we could make enormous progress in determining what factors are associated with homicides by the mentally ill (even though the problem of prediction will remain intractable because of its rarity). For example, if there is no difference between cases and controls in terms of CPA adherence, it would make no sense to criticise a team for not meeting every CPA criterion.

We continue to have inquiries routinely after each homicide by a person in contact, or recently in contact, with mental health services. It matters not if the person was judged by the service as not mentally ill! Does anyone really expect the next inquiry to carry a revelation, to discover something fundamental about what we have been doing wrong, which if corrected will prevent homicides in the future?

Worst-outcome cases

A homicide represents a worst outcome case for the professionals and service. Even if there were demonstrable failures on the part of the treatment team, does any professional deserve to be judged on the basis of what would be probably their worst ever case? As I sat during the press conference listening to an eminent, clearly very humane QC giving what could only be called a lawyerly account of ‘failures’ in the care of the patient in our inquiry, I wondered how she and her colleagues might have felt if the arrangements were slightly different. What if their worst case, perhaps a failure in the prosecution of a dangerous (not mentally ill) defendant had been the subject of an independent inquiry following a vicious homicide soon after acquittal. What if the psychiatrist chairman of the inquiry panel were giving an account at a press conference of a report, years in the preparation, to be presented to the Bar Council of, for example, evidence not adequately uncovered, poor communication between members of the prosecution team, arguments poorly presented, and so on, all in the presence of the victim's family and the sensationalist-seeking gaze of the media? Would this be reasonable?

A paradox: if homicides are preventable by a service, and are rare, the service must be good

There is also a strange paradox in a homicide inquiry. Inquiries are based on the assumption that a good service can prevent such events. If not, then it is presumably pointless. If so, 17 years of practice by a consultant's team, as in our inquiry, without another event like this, must indicate a high level of successful practice, especially if the service is forensic, dealing predominantly with dangerous patients. This makes it highly unlikely that systems failures will be associated with this particular event; systems failures are usually associated with a series of adverse events. A single homicide in such a context is thus likely to be one-off, or random, or largely patient- or circumstance-related, or represent an uncharacteristic lapse of judgement by an individual. In such circumstances an inquiry will not be informative about systems whose excellent outcomes in the past indicate their effectiveness!

It might be retorted: yes, homicides are rare; but the system must function in a constant state of preparedness so that such tragedies will never occur. For this to be a reasonable proposition it must be shown that: (a) the event is predictable (we have seen it is not); (b) there is a connection between the systems and the event (no evidence yet; a case—control study might reveal it); and (c) even if (b) is valid, that it is cost-effective to gear up the system to be ever vigilant for rare, unpredictable events. There are substantial costs to such a position, as we will see below.