Independent Association of Chronic Smoking and Abstinence With Suicide

Does smoking cause not only disease-related deaths but also suicide? Several prospective studies have shown an association between smoking and suicide (1–4). Because completed suicides are relatively infrequent, the relation has been studied with respect to suicidal thoughts, plans, and attempts—collectively known as suicide-related outcomes. Although positive associations have been found (5–8), the evidence is inconsistent. With completed suicides, a dose-response relationship with smoking was found (3,9), but no relationship was found when suicide attempt was the outcome (10). A dose-response relationship is typically used as a basis for inferring a causal relationship (11). When the analysis controlled for confounders, the association vanished in some studies (7,12,13) but remained significant in others (5,14,15). It has been argued that the smoking-suicide link is an artifact and reflects unmeasured common causes, such as mental illness and alcoholism (16,17). The case for smoking as an independent risk factor for suicide is supported by plausible physiological mechanisms.

Several smoking-related parameters have now been studied, including lifetime smoking duration (10,18,19), smoking recency (20), and cessation (21). Investigators studying lifetime smoking and subsequent suicidality have reached different conclusions. In one study, early nicotine dependence predicted subsequent suicidal plans, and this association was the same whether dependence remitted or not (10). Kessler and colleagues (10) argued that this evidence makes a causal interpretation problematic; instead these authors favored the “common cause” position. By contrast, a four-year prospective study reported that prior smoking predicted suicidality among patients with bipolar disorder, whereas suicidal ideation did not predict subsequent smoking or nicotine dependence (5). Because smoking preceded mental illness in this case, the authors favored an independent-risk interpretation of smoking.

Regarding smoking cessation, nicotine-dependent individuals who abstained from smoking in the past year were less likely to attempt suicide than those who continued to be nicotine dependent (22). Covey and colleagues (23) reported that individuals who had abstained from smoking for at least four years had a lower risk of suicide-related outcomes than did current smokers, whereas those who had abstained from 12 months to four years did not have a lower risk. On the basis of our literature review, we examined whether lifetime smoking duration, ever quitting smoking, and abstinence duration predicted suicide compared with other violent deaths.

An important way in which epidemiologic studies clarify this question is by using a comparison group that shares personal and social characteristics with suicide decedents. These shared characteristics, if present, would help isolate the contribution of smoking to suicide risk. Smoking has been studied in relation to accidental death (13,24,25), because, like suicide decedents, accident decedents tend to be aggressive and impulsive (1,26) and to have high rates of comorbid substance use disorders and mental disorders (27). It has been proposed that accidents and suicides arise from a continuum of self-destructive behaviors and that the behaviors differ in levels of intent (28,29). These self-destructive behaviors include tattoos, recklessness, and health-compromising behaviors (30). In case-control studies, it is important to match participant characteristics. Therefore, several case-control studies have used accident decedents as controls and suicide decedents as cases (31–34). Although not as common, suicide decedents have also been compared with homicide decedents, and the following shared risks were identified: access to handguns (35) and use of alcohol and illicit drugs in the home (36,37). Thus homicide decedents share with suicide decedents some environmental factors that predispose individuals to a violent death. Our review of the literature suggested that using homicide decedents and accident decedents as a control group is reasonable because both types of death are untimely and because levels of distress among kin would be comparable (38,39). We hypothesized that if smoking is independently associated with suicide, there would be an excess of suicide deaths among those with longer lifetime smoking, among smokers at death, and among more recent quitters after we controlled for confounding variables.

Methods

Sample

The data were from the 1993 U.S. National Mortality Followback Survey (NMFS) (40). The NMFS data set was previously used to study risk factors for suicide, but smoking was not examined in these studies (32,41). The survey involved a nationally representative sample of 22,957 individuals age 15 and older who died in 1993. In 1998, data from coroners’ reports on cause of death were added to the 1993 data. This variable included the following categories: natural, accidental, suicide, and homicide deaths. The kappa coefficient for manner of death, a measure of reliability across abstractors, was .89, which is very satisfactory (42). Following the psychological autopsy method, the decedents’ next of kin provided information on smoking and alcohol consumption, along with socioeconomic information. Although psychological autopsy has several limitations, it was reported that findings from psychological autopsies compare favorably with clinician interviews for patients with axis I and axis II disorders (43).

Our inclusion criterion was an affirmative answer to whether the individual ever smoked 100 cigarettes in his or her lifetime (N=10,010 of 22,957, 44%). Because the sample was selected to represent two racial groups (white and black), we excluded members of other racial groups. Also excluded were individuals who died from natural causes (N=12,875, 57%) and from undetermined causes (N=207, <1%). We used accident decedents (N=2,409, 10%) and homicide decedents (N=716, 3%) together as the control group and 989 suicide decedents (4%) as the case group.

Measures

Independent variables.

Smoking was assessed by three predictor variables: lifetime smoking duration, ever quit smoking, and duration of abstinence from smoking. Lifetime smoking was assessed by the single question, “How long did the decedent smoke cigarettes fairly regularly?” Smoking duration was treated either as a continuous variable (in years) or an ordinal one with four categories: zero to ten, 11–20, 21–40, and ≥41 years. As in a previous NMFS study (44), cessation of smoking was assessed by whether the decedent stopped smoking for at least three years before death and did not start again. Decedents who were reported to have abstained for some period were classified as former smokers or current smokers at death. Our third predictor variable, which indexed abstinence duration, also had continuous and ordinal versions. The continuous version was the number of days before death that the decedent had stopped smoking. For the ordinal version, we used the following categories: less than five years, five to less than 11 years, and 11 or more years.

Confounding variables.

Because depressive disorders and alcohol consumption are known to be associated with smoking, we statistically controlled for these confounders. Depressive symptoms were assessed by ten questions answered on a Likert scale, with the categories often, sometimes, rarely, and never coded in descending order from 4 to 1. These questions included how often had the decedent said things such as “I’m no good,” cried for long periods, had trouble sleeping, ate more or less than usual, had trouble concentrating, and talked about taking his or her life. A composite depression score variable was created by summing the responses; possible scores ranged from 10 to 40. Drug use in the last year of life was assessed by four questions about whether the decedent took heroin, marijuana, cocaine, or hallucinogens. A binary variable was created and coded as 1 if the decedent used at least one of these drugs and 0 otherwise. We reasoned that behaviors in the last year of life, including living alone, drinking alcohol, and keeping a firearm, are potential risks for a violent death, and thus we controlled for these variables. Veteran status was obtained from the death certificate and entered as a covariate. Living alone was assessed by a negative answer to the question, “Did the decedent live with somebody for at least three months in the last year of life?” Drinking in the last year of life was assessed by the question, “Did the decedent have at least 12 drinks of any kind of alcoholic beverage in the last year of life?” Keeping a firearm in the house was a strong predictor of suicide in a previous study (41). Demographic variables such as age, gender, and marital status were also used in the analysis.

Statistical Analysis

In univariate analyses, we compared age-standardized lifetime smoking and abstinence between cases and controls. These age-standardized variables were calculated by dividing lifetime smoking (or abstinence) duration by age. The quotient can be interpreted as the fraction of the lifetime that a given individual was a smoker (or abstained from smoking). We compared the proportion of individuals who ever quit smoking between the two groups. Survey-based Wald tests were used to calculate the difference in mean durations and proportions.

In multivariate analyses, survey-weighted logistic regression with stratification by geographical unit was used to model risk of suicide. Because suicide risk is not the same by gender, we developed separate models for males and females. A separate logistic regression model for each smoking predictor was fitted to avoid collinearity. Eight potential confounding variables were entered as covariates in the regression models: living alone, race, veteran status, age, drug use, depressive symptoms, drinking, and keeping a firearm in the house. All analyses utilized probability weights, and Taylor linearization was used for variance estimation as implemented in Stata, version 12 (45). Significance was set at a two-sided alpha of .05.

As a secondary analysis of existing data, and with no human subjects involved, institutional review was not necessary for this study.

Results

Among accident, suicide, and homicide decedents, the proportion of males was larger than the proportion of females. Among suicide decedents, the proportion of males was higher than it was in the combined control group (homicide and accident decedents): 84% versus 75%. Suicide and accident decedents were typically older (in their late 40s) than homicide decedents (early 30s), but the combined control group did not differ in age from suicide decedents. Compared with the combined control group, suicide decedents were more likely to be veterans, to have lived alone in the last year of life, to have a greater number of depressive symptoms, and to have kept a firearm in the house. Drinking in the last year of life and lifetime drug use did not differ between groups (Table 1).

Our univariate analysis showed that suicide decedents smoked for a larger fraction of their lives compared with decedents in the control group (F=37.91, df=1 and 12,424, p<.001). Suicide decedents were less likely to quit smoking in their lifetime than decedents in the control group (F=21.19, df=1 and 13,070, p<.001). Finally, compared with decedents in the control group, suicide decedents abstained for a smaller fraction of their lives (F=19.91, df=1 and 13,070, p<.001) (Table 2).

Our three multivariate models, which controlled for the eight potential confounders noted above, indicated significant associations between lifetime smoking patterns and suicide among males but none among females (Table 3). Quitting smoking and longer duration (≥11 years versus <5 years) of abstinence predicted lower odds of suicide. However, having a longer duration of lifetime smoking (≥41 versus ≤10 years) was associated with higher odds of suicide. For both lifetime smoking duration and abstinence duration, there was a significant linear trend among males but none among females.

Discussion

The objective of this study was to examine whether smoking is independently associated with suicide. The main findings are that male suicide decedents smoked for a larger fraction of their lives, were less likely to have quit smoking, and had a shorter abstinence duration than their counterparts who died as a result of homicide or an accident. No associations were found for females. Therefore, only the findings for males are discussed here.

The interpretation of our results follows the logic of accounting for known risk factors for suicide. If, having accounted for other confounders, smoking remains a significant predictor, then the finding constitutes evidence in favor of independent association. Otherwise, smoking would be an epiphenomenon—a risk that is already accounted for by the other variables. After all three models adjusted for commonly known risk factors, smoking remained a significant predictor. This implies that smoking is independently associated with suicide. However, this interpretation needs to be tempered by the fact that our measures were imperfect and the study had a retrospective design. Thus a causal relationship cannot be inferred, which is discussed below as part of the study limitations.

The dose-response association of suicide risk with lifetime smoking duration implies that the association with suicide is possibly causal in nature. Our result stands in contrast to a large prospective study that concluded that higher suicide risk among heavy smokers was entirely explained by other risk factors, notably heavy alcohol consumption and poor mental health (12). In our study, the association of chronic smoking remained significant after we accounted for covariates. Thus our results are more consistent with the interpretation that smoking introduces a physical or psychological toxin (46) into the nervous system. Chronic nicotine exposure causes a reduction in serotonin metabolites (47,48). The frontal cortex can also become less receptive to serotonin (49). Psychologically, smoking can worsen factors that predispose to suicide, such as increased mood lability, irritability, and depressed feelings, as well as behavioral tendencies such as aggression and impulsivity, which are known predictors of suicide (50,51). When individuals quit smoking, mood improves over time (51), and suicidality might decrease.

Early smoking initiation predicted suicide-related outcomes in other studies (10,18,19,50). Smoking initiation in childhood predicted later suicidal ideation and attempts for both males and females (18). Bolton and Robinson (52) estimated the population-attributable fraction (PAF) of nicotine dependence for suicide attempts at 8.4%. This was the third-highest-ranking disorder after major depression and borderline personality disorder. With that study’s cross-sectional design, the PAF estimate cannot be the basis for a causal claim, which was emphasized by the authors. In the context of previous findings, our study strengthens the case for smoking as a candidate risk factor for suicide. If smoking were truly an independent risk factor, then cessation should decrease risk (because the toxic effects would stop), but if it is an artifact, then smoking cessation should have no effect on risk (46). Our results regarding smoking cessation and abstinence duration are more consistent with the former proposition. The evidence on smoking cessation is mixed. Previously, a meta-analysis by Leistikow and Shipley (21) of three randomized controlled trials reported a nonsignificant difference in suicide rates between smoking cessation and control groups. Similarly, Kessler and colleagues (10) reported that early-onset nicotine dependence predicted subsequent suicide plans whether the dependence remitted or not. By contrast, Yaworski and colleagues (22) reported that nicotine-dependent individuals who abstained from smoking in the past year were less likely than smokers to attempt suicide. Berlin and colleagues (53) also studied smokers who had past major depression and who were attempting to quit. Those who were unsuccessful in abstaining had higher depression and suicide ideation scores than those who succeeded. The improvement of mood with abstinence (51) and the difference in depression levels between those who are successful and unsuccessful in quitting (52) support a causal interpretation. The discrepancy in reported findings could result from the lack of reliable measures. For example, one criterion for nicotine dependence in DSM-IV, the “persistent desire to cut down,” was criticized as being too subjective (54). In addition, smoking cessation activities fall along a spectrum, and all the studies did not use the same measure.

The interpretation of our results is subject to several important limitations. First, the psychological autopsy method relies on next-of-kin reporting, which is susceptible to hindsight bias and interpretation. Reports of depressive symptoms, suicidal ideation, and smoking behavior passed through the cognitive filters of the next of kin. Our exposure variables and covariates are likely to include some measurement error. Second, neither clinically validated diagnoses nor personality scores from the decedent were available. These would have shed light on whether smoking was related to latent disorders. Third, the association of smoking and suicide was not found among females. This might result from the gender difference in suicide risk or because of reporter bias. There is a “gender paradox” in that suicidal ideation and unsuccessful attempts are higher among females, whereas completed suicides are higher among males (55). The other possible explanation for the null finding among females is reporter bias. It is possible that next-of-kin reporters give a more favorable account of a female decedent than of a male decedent. Fourth, to address cessation and duration of abstinence from smoking, we restricted the sample to the subpopulation of individuals who had ever smoked. Fifth, since the exposures and outcomes of interest were assessed retrospectively, we cannot establish the onset of smoking vis-à-vis the onset of other risk factors, such as poor mental health and alcohol use. Studies with a prospective design that examine early smoking initiation or smoking cessation programs are better able to establish causation. Finally, it is possible that although we controlled for a large number of variables, some residual confounding remained. For these reasons, the independent association of smoking and suicide in this study cannot be used to infer causality.

Conclusions

Findings indicated a probable independent association between suicide and current smoking and longer lifetime smoking duration. The findings are additional grounds to investigate smoking as a possible independent cause of suicide.