Suicide among seniors is a significant health problem in north America, particularly for men in whom the rates rise steadily after 50 years of age.1 2 In contrast to younger suicides and parasuicides and attempters, seniors often choose more lethal means of suicide, give fewer warnings or indications of suicidal intent, and may more frequently engage in careful planning of the suicide as opposed to more spontaneous behaviour and actions.3 Seniors often live alone, which increases the likelihood of an uninterrupted act of suicide. Furthermore, medical co-morbidity and low physiological reserve may negatively impact the chance of recovery from a parasuicide event.

There remains a great deal that is unknown about elder suicides. For example, there have been relatively few controlled studies examining the relationship between physical illness and suicide risk in seniors.4^–11 Of these, six had small sample sizes,4^–9 six relied on proxy or self-report information4 5 7^–9 and only three were population based.6 10 11 Moreover, there have been even fewer controlled studies examining prescription medication use and suicide in older individuals. Toxicology studies have, however, indicated that antidepressants, opiates and benzodiazepines were often consumed before suicide either alone or in combination.12^–16 One population-based controlled study found that suicide attempts were related to benzodiazepine use and/or antidepressant use and that young men were at the greatest risk,17 whereas another controlled study of seniors has related the initiation of selective serotonin reuptake inhibitors to suicide in elderly men.18 Clearly, there is a need for further investigation to confirm these findings and to identify other medical, psychiatric and medication factors that may be related to suicide in this high-risk population. Of specific interest are several types and strengths of benzodiazepines that are not recommended for use in older individuals.19 20

The goal of this study was to examine elder suicides from a large population-based database using case–control methods to determine disease and medication factors related to suicide.

METHODS

Design

A case–control study was conducted using age and sex-matched controls in a 1 : 5 case to control ratio. The study compared medication use by British Columbia (BC) residents aged 66 years and older who had died by suicide (cases) with that of BC residents aged 66 years and older who had not died by suicide over a 10-year period (1993–2002). Complete prescription medication histories were available for this population. At age 66 years, each subject would have at least one year of insured medication use.

Data sources

Dependent and independent variables were derived for the study population from four databases maintained by the province of BC.

Canadian institute of health information hospital separation abstracts

These abstracts include International Classification of Diseases (ICD) version 9 diagnostic codes (up to 16), ICD-9 external cause of injury (E codes), age, sex, admission and discharge dates. One hundred per cent of hospitalisations in BC are captured through this system.

Physician claim file

This file contains physician claims for reimbursement from the BC Ministry of Health. Each claim contains a single diagnosis based on three-digit truncated ICD-9 diagnostic codes. Approximately 95% of physician–patient episodes of care in BC are reimbursed through the BC Ministry of Health. Physician claims files have been found to have a high concordance with chart reviews and patient interviews.21^–23

BC pharmacare file

This file contains the date of purchase, types of medications purchased, and the quantity of medication purchased by individuals who are eligible for reimbursement under the BC PharmaCare Plan. This includes all individuals 65 years of age and older. Although the accuracy of this file has not been formally established, these files exist for pharmacy billing and claimant reimbursement purposes and we assume that they are reasonably representative of the medication purchases of claimants. In addition, these data have been used in a number of medication use and injury studies in Canada suggesting acceptance by the research community as valid data sources.17 24 25

BC health insurance registration file

This file includes demographic information for each individual registered with the BC Ministry of Health. Health insurance coverage is universal in Canadian provinces.

BC Vital Statistics

This file contains the date and causes of death for BC residents. ICD-9 (1993–2000) and ICD–10 (2001–2002) external cause codes are assigned to any injury-related deaths.

Study population

Cases

A case was defined as a BC resident aged 66 years or older who was identified through BC Vital Statistics with the external cause of death coded as E950.x to E959.x for ICD-9 or X60 to X84 for ICD-10.

Controls

Controls were randomly selected from BC residents listed in the BC Health Insurance Registration file aged 66 years and older by sex and in five-year age groupings. Controls were subject to replacement if they had died from any cause before the index date of death for the matched case.

Matching criteria

Controls were matched to cases by five-year age groupings and sex. For comparison, a matched suicide event date was applied to each matched control subject as an index date for the retrospective medication and health service utilisation data. This controlled for any potential seasonal variation in medication use or health service utilisation. Matching was also conducted to control for a potentially unbalanced distribution of controls by age and sex that would occur if the simple random selection of cases had been conducted.26 That is, most of the cases are older men and most of the controls would have been younger and female. As sex and age are both strongly related to the exposure variables, medication use and co-morbidity, matching on these variables would improve the precision of any estimates resulting from the analysis.26 Controls were rejected without replacement if duplicates were selected.

Data linkage

Probabilistic linking was conducted by the Centre for Health Services and Policy Research at the University of British Columbia. The Centre for Health Services and Policy Research maintains a data warehouse of all administrative health data collected in BC.

Drug classification and other confounders

Demographic information available for analysis included the socioeconomic income quintile for the census tract of residence, whether the individual resided in the community or in supported care, and the geographical health region of residence. There are five regional health authorities in BC.

Medications were coded based on the American Hospital Formulary Service categories. The major classes of medications included in the study were: antihistamines, narcotic pain killers, anti-inflammatory drugs, antidepressants, barbiturates, benzodiazepines, tranquillisers, sedatives, cardiac drugs, antihypertensive drugs, diuretics, corticosteroids, diabetes agents, vasodilators and anti-lipid agents. Medication use for each case and control was quantified for the 30-day period before the suicide date. The total number of medications was the aggregate of the number of prescriptions filled, not the number of prescriptions from different physicians.

The co-morbidities included in the analyses were: liver disease, incontinence, osteoarthritis, dementia, depression/psychosis, neurosis, osteoporosis, renal disease, liver disease, alcoholism, hypertension, heart disease, diabetes, cancer, stroke, asthma and chronic obstructive pulmonary disease. Based on previous research into the relationship of hospitalisation with suicide in seniors11 each co-morbidity was given two levels (hospital or community), dependent on where the disease was coded. Hospital co-morbidity was coded for each individual with a hospitalisation record showing specific co-morbidities as primary or contributing diagnoses related to a hospital admission in the year before suicide. Community co-morbidity was coded when an individual had at least two episodes of care recorded for specific co-morbidities in the Physician Claim File over a two-year period before suicide. These co-morbidities were aggregated into a single three-level variable: no pre-existing co-morbidity (reference category); community-coded co-morbidity; and hospital-coded co-morbidity. Physician and hospital services received the same day as the suicide were not tabulated as they may have been part of the main outcome episode.

Data analyses

Descriptive statistics were generated using means, proportions and frequency distributions. Univariate conditional logistic regression analyses were used to identify independent variables significantly associated with the dependent variable (suicide) and are reported as odds ratios (OR) with 95% confidence intervals (CI) for three groups of independent variables: demographics, co-morbidities and medications. For each group of independent variables, multivariable conditional logistic regression was used to determine variables associated with the suicide outcome. A multivariable model was then developed that examined demographics, co-morbidities and medications from each variable group simultaneously to quantify their independent importance in relation to suicide. As many of the dependent variables could potentially exhibit strong correlations with each other, the assumption of multicollinearity was tested using methods recommended for logistic regression.27 The significance level was set at p<0.05 and analyses were conducted using LogXact version 7 (Cytel Software Corporation, Cambridge, Massachusetts, USA).

Finally, two further multivariable logistic regression analyses were conducted. First, the relationship between the number of prescriptions issued by all providers in the 30 days before suicide for narcotic pain killers and benzodiazepines and suicide was examined. Second, an examination of strong versus mild pain killers and recommended versus non-recommended benzodiazepines was performed. Strong pain killers included fentanyl, hydromorphone and morphine. Non-recommended benzodiazepines were categorised according to the latest Beers Explicit Criteria,20 which included chlordiazepoxide, diazepam, quazepam, halazepam, chlorazepate, flurazepam, lorazepam >3 mg, oxazepam >60 mg, alprazolam >2 mg, temazepam >15 mg and triazolam >0.25 mg. Tests for linear trends in the logistic models were conducted using the approach described by Selvin.28 Each multivariable model was also subjected to re-analysis stratified by sex to determine if sex was a confounder.29

RESULTS

Sample and suicide rates

Over the 10-year study period there was a total of 602 suicides by individuals aged 66 years and older in the province of BC with an aggregate rate of 13.2/100 000 per year (95% CI 13.1 to 13.3). The rate for men (21.9/100 000 per year; 95% CI 21.7 to 22.1) was significantly greater than the rate for women (6.5/100 000 per year; 95% CI 6.4 to 6.6). The use of a firearm was the predominant mechanism (28%) followed by hanging/suffocation (25%) and self-poisoning (21%) (table 1). There was a notable difference in the distribution of the mechanism of suicide between men and women, with firearms the most common cause for men (38%) and self-poisoning the most common cause for women (40%).

A total of 2999 age and sex-matched controls were randomly selected from BC residents. For 11 cases, controls were selected that had already been selected for previous cases. These duplicate controls were not used in the analysis. No selected controls had died during the data collection period of their matched case. The final case–control ratio was 1 : 4.98. As a result of matching, there were no significant age or sex differences between cases and controls (table 2).

Demographics and suicide

Table 2 also shows the relationship of demographic variables to suicide. In the univariate and group adjusted analyses socioeconomic quintile, community residence, and regional health authority were related to suicide. In the fully adjusted multivariable model that controlled for characteristics from all three variable groups, however, only socioeconomic quintile made a contribution to the prediction of suicide. The highest two income categories provided a protective effect in comparison with the lowest income category.

Co-morbidities and suicide

Table 3 characterises the co-morbidities related to suicide. At the univariate level, many different and varied illnesses were related to suicide. In the group multivariable model, depression/psychosis, neurosis, stroke, cancer, incontinence, liver disease and a previous history of self-harm were risk factors for suicide, whereas a diagnosis of dementia was protective. In the fully adjusted multivariable model all of the variables in the co-morbidity group multivariable model remained significant with the exception of dementia.

Medication use and suicide

The relationship between medication use and suicide is demonstrated in table 4. Antidepressant, benzodiazepine, barbiturate, tranquilliser and narcotic pain killer use were all related to suicide. In the medication group adjusted analysis antidepressant, benzodiazepine, barbiturate, and narcotic pain killer use were all related to suicide, whereas diuretic use was significantly related to a lower likelihood of suicide. In the fully adjusted multivariable model benzodiazepine and narcotic pain killers remained significant risk factors for suicide with the use of diuretics exhibiting the opposite effect. An examination of the variance inflation factor did not yield any values greater than 2.5, indicating that multicollinearity between independent variables was not affecting the results.27

To explore further the relationship between medication use and suicide, benzodiazepine and narcotic pain killer treatments were entered into an adjusted logistic regression model to determine if there was increased risk with multiple prescriptions. A dose–response relationship for both drugs is shown in column one of table 5. A similar relationship was also exhibited when poisoning-related suicides (n  =  128) were removed from the analysis (column 3, table 5). There were significant linear trends across the multiple medication categories for pain killers and benzodiazepines.

Finally, using an adjusted logistic regression model, the relationships between the strength of pain killers as well as benzodiazepine type (appropriate or inappropriate for seniors) was examined (table 6). This analysis showed that suicide was more likely to occur for individuals prescribed strong pain medications or inappropriate benzodiazepines. This relationship was the same for both poisoning and non-poisoning suicides. For pain medication, the odds ratios for strong pain medication were significantly greater than those for mild pain medication. For the benzodiazepine classes, the odds ratios for benzodiazepines not recommended for seniors were not significantly greater than those for recommended benzodiazepines. There were, however, significant linear trends across the benzodiazepine categories. Stratified analysis by sex did not alter the relationships identified in the matched analyses.

DISCUSSION

This large population-based study has provided important information regarding senior suicides. First, it is consistent with previous work that has identified significant associations between suicide in older individuals and medical or psychiatric diseases. Second, these results have identified a relationship between the use of benzodiazepines and suicide that has not previously been reported. In particular, there was an increased risk of suicide when non-recommended benzodiazepines were prescribed. This relationship was maintained when poisoning suicides were removed from the analysis. Finally, those individuals with severe pain appear to be at increased risk of suicide. Clearly, the results are important to patients and their families, healthcare providers and policy makers caring for older individuals.

Toxicological studies have found psychotropic drug levels from trace to lethal in retrospective analyses of completed suicides.12^–16 Few controlled studies have, however, been conducted to examine whether these drugs are possible risk factors for suicide. Neutel and Patten17 found that antidepressants and benzodiazepines were related to suicide attempts in a population-based controlled study. That study did not examine specific types or dosages of benzodiazepines and the authors acknowledged that confounding by indication was the likely explanation for the relationship they found between antidepressants and suicide. In the current study the analysis was extended by creating subgroups of prescribed benzodiazepines; appropriate and inappropriate prescriptions on the basis of Beers Explicit Criteria.20 The inappropriate group of benzodiazepines included those agents with prolonged half-lives and those in which the dosage may exceed what is considered safe for elderly patients. Simply put, inappropriate benzodiazepines are either exceeding dose recommendations or persist too long in their system for the physiological capacity of the elderly. For the long-acting benzodiazepines this can result in sedation that can last up to several days and strong benzodiazepines should be prescribed in lower dosages as a result of increased sensitivity in older patients.20

Benzodiazepines are central nervous system depressants and may exacerbate undiagnosed depression in seniors. This effect may be intensified for non-recommended benzodiazepines, which linger systemically or are beyond advised dosages. It has also been hypothesised that the use of benzodiazepines may lead to uninhibited and poor impulse control leading to suicidal impulses in some individuals.30^–32 Although there is limited empirical evidence to support this hypothesis, it is possible that non-recommended benzodiazepines compound this potential effect. These results also suggest that individuals filling multiple prescriptions of benzodiazepines were at greater risk, indicating that heavy and chronic usage of this class of drugs is a marker for depression or may be a mitigating factor leading to suicide through poor impulse control. Doctors prescribing these drugs may not have diagnosed underlying depression or may be using benzodiazepines as a primary treatment of depression rather than an adjunct therapy to true antidepressant medications.

Physical disease was identified as a strong predictor of suicide in this study. There also appeared to be a gradient in which those hospitalised as opposed to being treated in the community were at greater risk of suicide. The association between recent hospitalisation and suicide in seniors has been reported by others.11 The specific relationship between cancer and suicide found in our study has been reported previously.33^–35 To our knowledge, however, the association between liver disease and suicide has not been quantified empirically. The relationship between liver disease and suicide in this study was large (OR 3.08; table 4). Liver disease has been related to depressed mood in patients with serious diseases such as hepatitis C and cirrhosis.36 37 Long-term liver disease is associated with greater levels of depression,38 as well as the fatigue caused by liver disease.39 Recent findings have also determined a connection between non-alcoholic steatohepatitis and depression.40 In our population, very few cases of liver disease were coded as being related to hepatitis C or alcohol abuse. Others have found a relationship between stroke and suicide as well as genitourinary disease and suicide.33 41 It has been reported that approximately one third of individuals with stroke have depression.42 Those diagnosed with incontinence have also been found to have greater levels of depression when compared with healthy controls.43 Depression may also overlap with agitation and insomnia in those patients with medical co-morbidities.44 It is likely that many of the suicides in our study were individuals with undiagnosed or undertreated depression related to their medical co-morbidity.

A concurrent psychiatric diagnosis is commonly associated with suicide, and our findings support this relationship. Both depression/psychosis and neurosis were associated with suicide outcome. There was also an observed gradient between community-diagnosed mental health problems and those contributing to a hospital episode, with the community diagnoses having a weaker relation to suicide. Moreover, a parasuicide in the previous year exhibited a very strong association with subsequent suicide (table 3). Clearly, seniors who are parasuicides are at an extremely high risk of becoming suicides and should be monitored very closely at this age. Twenty-eight per cent of parasuicides did not have a coded diagnosis of depression. As unrecognised depression is common in the elderly,45 46 and many depressed individuals fail to communicate their despair to their physicians before suicide,33 these findings have important implications for depression screening in the elderly. More importantly, only 15% of those with a history of parasuicide had filled a prescription for antidepressants in the month preceding suicide, indicating that undertreatment of depression is also a concern. Patient compliance may also be an issue if patients with depression are prescribed medication but choose not to fill the prescription. These severely depressed seniors would benefit from a follow-up protocol that encourages the use of the appropriate medication as well as other non-pharmaceutical therapies.

The risk gradient identified in this study indicates that individuals taking stronger pain medication were at a greater risk of suicide. Similar to severe disease, severe pain is probably related to increased levels of depression and hopelessness. Our findings are consistent with another population-based Canadian study using similar methods that found the same risk gradient from mild to strong pain medication.10 In addition, we also found that individuals filling multiple prescriptions of pain medication in a 30-day period were at greater risk of suicide, again supporting the theory that patients with pain are at greater risk. Counting the number of prescriptions filled does not imply a cumulative dose, but is more likely to be related to pain of a more chronic nature. The literature suggests that chronic pain is antecedent to depression.47 Clinicians should thus consider that those patients with chronic pain may also develop depression at some time during the course of treatment.

There are several limitations of this study that warrant discussion. First, we used medications purchased as a proxy indicator for medication use/non-use. It is possible that some medications were prescribed but not purchased and purchased and not subsequently taken. We were thus not able to determine the cumulative dose of any of the medication classes examined. The use of this proxy measure, however, is likely to dilute the strength of the observed associations rather than produce spurious results. Second, no measure of non-prescription medication use, complementary and alternative medication use, alcohol or illicit drug use was available from this database linkage study. Information on these variables is not available in any of the databases used in this study, therefore only the relationship of prescription medication use with suicide was available for analysis.

Third, there is a potential underreporting of medical conditions (co-morbidities) because we obtained this information from administrative data. The absence of a coded disease does not necessarily mean an absence of the condition. The measurement of these conditions also relies on accurate and consistent coding by medical professionals. If there is inconsistency in coding or diagnostic misclassification occurs between similar conditions, some error may have been introduced. Another issue that arises with this type of research is that multiple medication classes may be used to treat multiple co-morbid classifications (eg mental health or heart problems). This makes it particularly difficult to distinguish the combined or singular effects of any specific medication and co-morbidity pairing. The study was also retrospective in nature; however, alternative approaches such as prospective cohorts would be expensive, laborious and protracted given the low event rate of this condition. Finally, other known risk factors for suicide in the elderly such as social isolation, the recent loss of a spouse, or family history of suicide were not included in our analyses as these variables were not present in the administrative health databases.5 48 49

Notwithstanding these limitations, these results are methodologically rigorous, appear valid and should provide important information for multiple stakeholders. We used a population-based sample in which the control group was representative of the total population, which increases the external validity of the study. As healthcare coverage in Canada is universal, with every person over the age of 65 years having premium-free health insurance and medication coverage, the dataset includes the entire population of interest, therefore reducing the potential effects of selection bias. Cases and controls were temporally matched to eliminate seasonal variability, and identical measurements for all variables in both the case and control groups were obtained. All health data were collected in a uniform health system, which minimised measurement bias.

In conclusion, this study has confirmed the relationship between medical or psychiatric co-morbidity and suicide in seniors. In addition, new and potentially useful information confirms that certain types and dosages of benzodiazepines are harmful to seniors and should not be prescribed.