Abstract
People with a history of mental illness are at much greater risk of suicide than those without such a history [1]. Australia, like many other countries, has developed a policy framework for suicide reduction that involves targeting this high-risk group [2–4]. This policy framework suggests that mental health care providers have a key role to play in reducing suicide among this group by providing optimal care for all those with mental illness, and identifying and intervening with those who are most at risk. For the latter to occur, it is necessary to identify patient-based and treatment-based risk factors.
Numerous studies have considered patient-based characteristics associated with suicide among current or former psychiatric patients. Appleby [5] and Davis and Schreuder [6] reviewed such studies and found the following characteristics were common among patients who suicided: particular psychiatric diagnoses (e.g. psychotic, affective, substance use or personality disorders); other clinical factors (e.g. comorbidity, gradual onset of disorder or previous suicide attempts); and sociodemographic factors (e.g. being single, unemployed, male young or socially isolated).
There is also a body of studies that have considered treatment-based characteristics associated with suicide among psychiatric patients. For example, a recent largescale audit of suicides by psychiatric patients in Victoria found that half of these patients had received care in the month before death [7]. The results of a comparable study in the United Kingdom were even more striking, with 50% receiving care in the week before death [8]. These patterns of close contact are consistent with findings from elsewhere in the literature, as demonstrated in a recent review of 24 studies [9].
The difficulty with many of these studies is that they are purely descriptive, and involve no comparison groups. This is particularly so for studies that consider treatment-based characteristics, but also applies to many that examine patient-based factors. Without any comparison group, it is not possible to determine whether, for example, being single is a particular risk factor for suicide, or merely a common characteristic of all psychiatric patients. Likewise, it is not possible to determine whether particular patterns of care are unique in any way to psychiatric patients who attempt suicide.
Several recent studies have attempted to address this problem by using case-control designs. In Australia, Cantor et al. [10] identified 34 individuals who had suicided and had ever made contact with one of four community psychiatric services, and individually matched them (by age and sex) to 34 controls drawn from the same services. They found no difference between cases and controls in terms of socio-demographic or clinical factors. In the United Kingdom, Dennehy et al. [11] identified 70 suicide cases who had had a psychiatric admission in the 5 years prior to death, and individually matched them (by age, sex, diagnosis and date of admission) to controls identified by block randomisation of hospitals in the area. They found that cases and controls could not be distinguished on most socio-demographic, clinical and treatment-based variables (with the exceptions being communication of suicidal ideas and involuntary status). In an extension of this study which involved 149 cases and controls, Appleby et al. [8] again found that cases could not be differentiated from controls on the basis of most socio-demographic or clinical variables, but that the former were more likely to have their care reduced at their final community appointment.
The current study built on this work by considering a much larger group of cases and controls. In addition, it refined the case-control methodology. One of the key criteria for selecting controls is that they should have the potential to be included as cases if they exhibit the outcome of interest [12]. The selection process in the current study ensured that the controls represented not just all psychiatric patients, but psychiatric patients who would have been included as cases had they suicided. Evidence that each control was alive and resident in Victoria beyond the death date of his/her case was specifically sought. In addition, no explicit criteria were used to match cases to controls, on the grounds that this would limit the variables that could be examined in the analysis. For example, it was decided not to match on date of last admission, since this would have precluded opportunity to determine whether there was a difference between cases and controls in terms of time since last admission.
Method
Choice of study design
We conducted a matched case-control study to estimate the strength of the association between risk of suicide among psychiatric patients and a range of patient-specific (e.g. demographic and clinical characteristics) and treatment-specific variables (e.g. frequency, duration and recency of care) obtained from the Victorian Psychiatric Case Register (VPCR). We chose this design, rather than a complete cohort analysis, for reasons of efficiency (abstracting data for the entire cohort was not feasible) and to avoid confounding arising from long-term changes in treatment patterns or suicide risk. We required each control to be alive (i.e. at risk of suicide) and to have had the same opportunity to receive care up to the date of death of the corresponding case; in other words, matched cases and controls were in the same risk set.
Selection of cases
By linking the Victorian Department of Justice's Unnatural Deaths Register (VUDR) to the Patient Master Index (PMI) from the VPCR, we identified 597 cases who suicided between 1 July 1989 and 30 June 1994 and had at least one occasion of service recorded on the VPCR.
Selection of controls
To be in the same risk set as the case, each control had to be: (1) registered on the VPCR prior to the suicide date of the case; and (2) alive at or beyond the suicide date of the case (and therefore have the opportunity to receive care up to that date). For each case, a pool of potential controls was identified that comprised all patients who were registered on the VPCR before the case's suicide date. Apart from vital status up to the death date of the case, no specific matching criteria were used to select controls, since any variable used for matching could not then be analysed for its predictive value as a risk factor. Individuals were simply randomly drawn from this pool until one was identified who satisfied the criterion of being alive at or beyond the suicide date of the case. Individuals could demonstrate this in one of two ways. The first way involved their being registered on the current Victorian Electoral Roll (VER). Since the current VER post-dated the latest possible suicide date (i.e. 30 June 1994), any individual who was registered on the VER was taken to be alive beyond the suicide date of the case. The alternative way in which individuals could demonstrate their positive vital status was having service use recorded on the VPCR beyond the suicide date. As noted above, each selected control was allocated an index date that corresponded to the suicide date of the case. This date is referred to as the ‘suicide/index date’ for both cases and controls in the remainder of this paper.
Characteristics of the databases used in the study
The VPCR is part of Victoria's state-wide medical record system that registers patients and records activity associated with them. It is a person-based register that dates back to 1961 and records the contacts with all specialist public inpatient and community mental health services in Victoria (including adult, child and adolescent, psychogeriatric and forensic services). Any individual making contact with any of these services is assigned a unique identifier that allows him or her to be ‘tracked’ across services. It should be noted that the VPCR does not cover contacts with private sector services, such as private psychiatrists, general practitioners, private psychologists and private psychiatric hospitals. Individuals who made contact with specialist public sector services but whose normal residential address was outside Victoria were not considered in the current study.
The VUDR is a compilation of all investigations by the Victorian coroner into sudden and unexpected deaths. All deaths deemed suicides by the coroner were included in the audit, unless they were by persons not normally resident in Victoria. Individuals who were normally resident in Victoria, but who died by suicide outside the Victorian Department of Justice's jurisdiction were not captured in the VUDR.
The VER comprises information requested from electors when they enrol in Victorian electorates. It is regularly maintained for the purposes of conducting Federal and State Parliamentary elections, such that individuals who have died or moved interstate are removed from it.
Data extraction
Once each case-control pair was identified, data were extracted from the VPCR for all individuals.
The data extraction process yielded three types of data. Identifying information defined each individual by unit record number, case/control status, and the suicide/index date associated with their caseset. Patientbased variables included a range of socio-demographic and clinical variables. Treatment-based variables described patterns of care relative to the suicide/index date, in terms of frequency of particular types of care (e.g. total number of inpatient admissions), duration of particular types of care (e.g. total number of inpatient days) and recency of particular types of care (e.g. time since last inpatient day). The patientbased and treatment-based variables are summarised in Table 1.
Predictor variables
Data analysis
Conditional logistic regression was used to estimate the strength of association between the patient-based and treatment-based predictor variables and the risk of suicide. Specifically, an iterative model-building process was used. The objective was to determine which variables were independently predictive of suicide, after adjustment for other variables. At each stage in this process, the effect of a given variable was considered; if the variable was not significant in its own right or did not substantially influence the coefficient of (i.e. confound) another variable, it was excluded from the model. Data were analysed using Stata (Version 6) [13].
Results
Table 2 presents crude and adjusted odds ratios for those variables found to influence the risk of suicide among psychiatric patients. Two patient-based factors were associated with increased risk. Compared with psychiatric patients who were not in the labour force, those who were employed, unemployed or of unknown employment status were significantly more likely to suicide, and male psychiatric patients were more than twice as likely to suicide than female patients. Three treatment-based factors that were associated with recency of contact influenced risk. Compared with those who had no inpatient admissions (i.e. had been treated in the community only), those who received recent inpatient care were more likely to suicide, with the risk peaking for those who had received their last inpatient care in the period 1 week to 1 month prior to the suicide/index date. In comparison with patients who received no community-based care, those who received recent community care were at greater risk of suicide. The risk was greatest in the period up to 7 days before the suicide/index date, and dissipated over time. Those whose most recent occasion of service was a registration only with no subsequent recorded service activity (as opposed to community-based assessment or treatment) were at increased risk of suicide. A registration only occurred for a variety of reasons, including that the appointment was for an assessment, no further treatment was offered, the patient was offered the option of making an appointment at a later date, or treatment was offered but the patient did not return for a further appointment.
Conditional logistic regression model for suicide
Cases could not be distinguished from controls on the majority of patient-based factors: age; country of birth; marital status; rural/urban area of residence; socio-economic disadvantage; or diagnosis. Likewise, there was no significant difference between cases and controls in terms of several treatment-based factors, particularly those related to duration and frequency: duration on the register; total number of admissions; and total number of community days.
One treatment-based factor is worthy of note. When treated as a continuous variable, the total number of inpatient days was positively and significant associated with suicide risk. However, when it was treated as a categorical variable, it became negatively associated with suicide risk, it only just retained statistical significance, and no clear pattern could be discerned. Difficulties arose because of the ‘structural zero’ created by it being impossible for those who had never been inpatients to have had any inpatient days recorded. For these reasons, total number of inpatient days was dropped from the final model.
Discussion
The current study is novel in several ways. The vast majority of other studies that have been done in the area have been descriptive, rather than analytical, and involved no comparison groups. They have also tended to focus on patient-based, rather than treatment-based, risk factors. The study improves on these by adopting a matched-pairs case-control methodology, and giving equal weight to patient-based and treatment-based characteristics. It also builds on the few existing case-control studies in the area [8, 10, 11], by its unique attention to ensuring that each case-control pair were in the same risk set. This permitted an examination of the patterns of service use up to the suicide/index date of each pair. Other studies have also considered service use data for case-control pairs up to the date on which the case suicided, but have not given consideration to whether the control had the same opportunity to receive care as the case. These studies have also used more stringent matching criteria (e.g. age, sex, diagnosis and admission date), thus precluding any possibility of considering the extent to which these factors might be predictive of suicide.
The present study identified two patient-based risk factors for suicide. Being male increased the risk of suicide; not being in the labour force was protective against it. The former finding is consistent with the results of many descriptive studies, and might be explained by the more violent methods typically used by males who attempt suicide, as compared with their female counterparts. The latter finding is more complicated. Other studies have indicated that not being in the labour force is protective against suicide, often suggesting that individuals outside the labour force may be parents with dependants who have been shown to be at relatively low risk of suicide because of their commitment to others. These studies have also tended to suggest that being unemployed is a key characteristic of patients who suicide. In the current study, while those who were unemployed were at increased risk, the risk for those who were employed was even more elevated. This clearly warrants further investigation.
Three treatment-based risk factors were found to be associated with suicide. Moderately recent contact with inpatient services placed patients at maximal risk, as did very recent contact with community care. So too did having a last occasion of service that was a registration only, as opposed to a community contact for assessment or treatment. Taken together, these findings suggest that there is something unique about the patterns of care that arise for psychiatric patients who suicide. They can perhaps be grouped into two types: those who are in receipt of inpatient care (presumably indicating that their symptomatology is relatively severe) and who use the period after discharge to plan their suicide attempt; and those who make contact with community-based services (often for the first time) as a ‘cry for help’. This suggests that there is potential for services to improve risk assessment for recent inpatients in the period immediately following hospital discharge, and for them to ensure that there is rapid follow-up of first appointments for at-risk community-based patients.
Perhaps even more striking than the above positive findings is the fact that cases and controls could not be distinguished on the majority of patient-based and treatment-based variables. This observation is consistent with the findings of the small number of other case-control studies that have been conducted in the area [8, 10, 11]. This suggests that previous descriptive studies have merely been reporting on factors that are common among all psychiatric patients, rather than describing characteristics that are unique to those who suicide. Take, for example, being single (i.e. never married or once married). In the current study, 431 cases (or 72%) fell into this category. In an earlier descriptive audit of these cases, it was reported that being single was a common characteristic of psychiatric patients who suicided [7]. However, supplementary analysis in the current study found that 397 controls (or 66%) were also single, indicating that this situation is commonplace among psychiatric patients as a general group. Marital status was eliminated from the conditional logistic regression model because of its lack of statistical significance.
Despite its methodolgical rigour, the study had several limitations that should be acknowledged:
Generalisability of findings
The study was restricted to specialist public sector psychiatric patients who died in (or remained alive and resident in) Victoria. One of the key criteria involved in making a decision to provide public sector care (particularly inpatient care) to patients in Victoria is that they have high levels of suicide risk. We are confident that our coverage of these patients was comprehensive, but acknowledge that the picture might be different for private sector psychiatric patients in Victoria, or for public sector patients in other geographical locations or jurisdictions, particularly ones that might use different criteria for providing care.
Selection bias
By necessity, noncomparable criteria were used to select cases and controls. Both were drawn from the universe of patients registered on the VPCR, but different means were used to determine their vital status. The vital status of each case was determined by their inclusion on the VUDR whereas the vital status of each control was determined by their presence on the VER or by their having activity recorded on the VPCR after a particular date. These noncomparable selection criteria may have introduced some bias. For example, it may be that some psychiatric patients do not appear on the electoral roll, despite being alive and resident in Victoria. These are likely to be those who are itinerant, or who have been resident in long-term inpatient facilities for a number of years. The selection process would have rendered it more likely that controls would have displayed these characteristics.
Measurement bias
The VPCR exists as a by-product of an administrative process that requires clinicians to submit patient-based and treatment-based details as a condition of funding of public sector mental health services. These data are entered by medical records staff, who are guided by a range of computerised validation rules. The strength of this dataset is that it provides comprehensive, population-based data on all individuals who come into contact with any public sector mental health service in Victoria, and the amount of missing data is small, but it is acknowledged that it is essentially a minimum dataset that provides relatively broad, high-level information on each patient, and it was not possible to validate the quality of any of the VPCR data against any sort of ‘gold standard’.
Confounding
We felt that it was appropriate to deal with confounding in the multivariate analysis, rather than in the design. There is the possibility, however, that other potential confounders were operating.
Conclusions
These limitations aside, certain conclusions can be drawn. It is difficult for clinicians to predict suicide. Not only is suicide a rare event, but patients who subsequently suicide ‘look’ similar to patients who do not. Despite this, it is clear from clinical experience that not all psychiatric patients have the same level of suicide risk. Prediction of suicide is the wrong concept; risk assessment makes more clinical sense. The corollary of this is that developing screening tools with sufficient sensitivity and specificity to be useful to clinicians is unlikely to be successful, and it may be more appropriate to concentrate on developing more general approaches aimed at ensuring that all psychiatric patients receive optimal care. Such care includes individualised assessment of suicide risk, and early and intensive follow-up of those deemed to be at high risk. This will involve providing appropriate support and training to those working in the specialist mental health sector.
Footnotes
Acknowledgements
This study is an extension of work made possible by funding from the Mental Health Branch of the then Victorian Department of Health and Community Services. Adam Clarke and Simon Palmer of Strategic Data assisted with the original identification of cases, and the identification of potential controls. Merilyn Yemm of the Victorian Electoral Commission was responsible for matching potential controls to the Victorian Electoral Roll, in order to identify actual controls.