Interaction Domains and Suicide: A Population-based Panel Study of Suicides in Stockholm, 1991-1999 Peter Hedström Liu Ka-Yuet Monica K. Nordvik Social Forces, Volume 87, Number 2, December 2008, pp. 713-740 (Article)

Published by The University of North Carolina Press DOI: 10.1353/sof.0.0130

For additional information about this article http://muse.jhu.edu/journals/sof/summary/v087/87.2.hedstrom.html

Access Provided by Columbia University at 08/31/10 11:00PM GMT

Interaction Domains and Suicide: A Population-based Panel Study of Suicides in Stockholm, 1991-1999 Peter Hedström, Singapore Management University and University of Oxford Ka-Yuet Liu, Columbia University Monica K. Nordvik, Stockholm University This article examines how suicides influence suicide risks of others within two interaction domains: the family and the workplace. A distinction is made between dyad-based social-interaction effects and degree-based exposure effects. A unique database including all individuals who ever lived in Stockholm during the 1990s is analyzed. For about 5.6 years on average, 1.2 million individuals are observed, and 1,116 of them commit suicide. Controlling for other risk factors, men exposed to a suicide in the family (at work) are 8.3 (3.5) times more likely to commit suicide than nonexposed men. The social-interaction effect thus is larger within the family domain; yet work-domain exposure is more important for the suicide rate because individuals are more often exposed to suicides of coworkers than family members. As Bearman (1991) noted, an important insight in Durkheim’s classic study of suicide was his recognition that the structure of social relations is important for explaining variations in suicide rates. These insights were only partly explored in Durkheim’s own work, however, and lack of relevant data means that we still know far too little about how the social structures in which individuals are embedded influence their suicide risks and the suicide rates in society at large. In this article we focus on one specific aspect of this: how suicides among those with whom a focal individual interacts influence his/her own suicide risk, and how such processes are conditioned by the social structures in which they take place. Although we are primarily interested in the structural aspects of suicides, we will pay a great deal of attention to individual-level processes. Such processes are essential for understanding why we observe what we observe, and when it comes to empirically This research has been supported by grants from The Swedish Council for Working Life and Social Research and the NEST/Path Finder initiative of the European Community, the DYSONET and MMCOMNET projects. We wish to thank Yvonne Åberg, Peter Bearman, Francois Collet, Fredrik Liljeros and the referees for their useful comments. Direct correspondence to Peter Hedström, School of Social Sciences, Singapore Management University, Singapore 178903, e-mail: [email protected], or Nuffield College, New Road, Oxford OX1 4RB, United Kingdom, e-mail: [email protected]. © The University of North Carolina Press

Social Forces 87(2), December 2008

714 • Social Forces 87(2)

assessing the importance of network effects it is essential to control for potentially confounding individual-level factors. Our approach thus differs from what in the suicide literature is often referred to as a “sociological” or “Durkheimian” approach to suicide research, an approach that focuses on associations between suicide rates in different population groups and various properties of these groups.1 Methodological developments during the past few decades as well as the increased availability of data sets with detailed information on the individuals at risk and their local social environments, allows for more precise analyses of the social-structural aspects of suicide. Rather than focus exclusively on macro-level patterns, we focus on the links between micro and macro, i.e., on how individuals’ suicide risks are affected by characteristics of their social environments, and on how the social domains in which suicides take place influence the extent to which others are exposed to and influenced by them. As will be discussed below, a range of studies suggest that exposure to the suicide of another person is likely to increase the suicide risk of the exposed individual. The data used have not always been ideal, however. Ecological studies have the shortcoming of not adequately controlling for individual-level risk factors, and in most ecological studies it is not known who has actually been exposed to the suicide of another person (Baron and Reiss 1985). With their small sample sizes and cross-sectional designs or short follow-up periods, many individual-level analyses lack the power to detect how suicides affect the suicide risks of others. We have access to a unique Swedish database that allows us to overcome some limitations of previous studies. The database is a panel including all adults who ever lived in the greater Stockholm metropolitan area during the 1990s. It includes detailed socio-demographic information on all these individuals, including causes of death for all who died during this period of time. In addition, and of crucial importance for this study, it contains information that allows us to identify co-workers, relatives and family members of a suicide. The article also makes a theoretical contribution by highlighting the importance of two types of network effects: a dyad-based social-interaction effect and a degree-based exposure effect. We show that focusing on one of these effects at the expense of the other can easily lead to a biased understanding of the role that networks play in explaining suicide rates and other macro-level outcomes. Network Effects Suicide risks have been linked to a range of medical and biological factors including various psychiatric disorders and personality characteristics (see

Interaction Domains and Suicide • 715

Figure1.1.Suicide-Exposure Suicide-Exposure Effects Effects from Froman anAction-Theory Action-theory Perspective Figure Perspective Beliefs of Individual j Suicide of Individual i

Desires of Individual j

Risk that Individual j Decides to Commit Suicide

Opportunities of Individual j

Jacobs, Brewer and Klein-Benheim 1999). Although such factors typically operate “behind the back” of the individuals, suicide researchers such as Shneidman (1969) have long suggested that suicide should be seen as a purposeful act aimed at solving one’s problems. It is useful to explicitly analyze suicides as intentional actions because by adopting an action perspective we are better positioned to understand the mechanisms through which interactions with others are likely to influence a focal individual’s suicide risk (see also Michel and Valach 1997). Actions have been conceptualized and analyzed in numerous ways, but a common denominator among most of these approaches is a view according to which actions are seen and analyzed as a joint product of reasons and opportunities.2 As suggested by Davidson (1980), Elster (1983) and others, the cause of an action can be seen as a specific constellation of desires, beliefs and opportunities in the light of which the action appears reasonable and understandable. Beliefs and desires are mental events that can be said to cause an action in the sense of providing reasons for the action; they have a motivational force that allows us to understand and, in this respect, explain the action. As illustrated in Figure 1, to the extent that the suicide of one person (“Individual i”) influences the suicide risk of another person (“Individual j”), this influence must be mediated via the beliefs, desires or opportunities of the latter person (see also Hedström 2005). In order to better understand why and when network effects are likely to be important, it is essential to specify clearly the mechanisms that are likely to be at work (see Hedström and Swedberg 1998 for a general argument to this effect). One belief-based interaction mechanism that has received some attention is concerned with how exposure to the suicide of another person may change the focal person’s beliefs about how others are likely to react were he to commit suicide. For example, if a suicide generates considerable sympathy towards the deceased individual, this may change the focal person’s beliefs about how others are likely to react to his suicide in such a way that he

716 • Social Forces 87(2)

becomes more prone to commit suicide (see Cutler, Glaeser and Norberg 2001; Hankoff 1961). The mechanism is exemplified by the remark made by a 13-year-old boy during the funeral of someone who committed suicide: “How nice it would be to have all those people crying and making a fuss over me.” (Hezel 1985:120)3 It is this mechanism that explains why most media guidelines suggest that “glorification” of suicide and sensational mass media reporting should be avoided (CDC 1994; WHO 2000). Suicides of others may also influence the focal person’s beliefs about his ability to commit suicide. As we know from research within numerous other areas of life, the availability of social models is one key source of self-efficacy (Bandura 1994). Seeing that people similar to ourselves succeed in performing a particular type of action tends to strengthen our beliefs in our own ability to perform the same type of action. According to Linehan et al. (1983), for example, such beliefs (what they refer to as “courage”) are important for explaining why some suicidal individuals, but not others, decide to take their own lives. Thus, the suicides of others may increase the focal individual’s suicide risk by changing the focal person’s self-efficacy beliefs. Desire-related interaction mechanisms also are likely to be of importance. In almost every culture, there is a strong normative pressure against committing suicide (Farberow 1975).This may dissuade some individuals from committing suicide who in another normative context would have done so. In most areas of life, such normative pressures tend to be density dependent in the sense that the strength of the normative pressure against doing X is a negative function of the number of other people already doing X (see Hedström 2005). This implies that an increased exposure to the suicides of others will reduce the normative pressure against committing suicide, thereby increasing the suicide risk. In addition to these belief- and desire-related interaction mechanisms, it also seems important to consider so-called “trigger mechanisms” or events that change the cognitive status of an action alternative from being a mere theoretical possibility to a consciously reflected upon alternative.4 As argued by Åberg (2003), each day of our lives we are faced with an almost infinite number of possible action alternatives, but we consciously reflect upon only some of them. In order to explain why individuals do what they do, it is often essential to take into consideration events that change the cognitive status of different action alternatives. Suicide is a theoretical possibility for any of us, but rarely is it a consciously reflected upon alternative (see Rudd 2000). The suicide of another person may act as a trigger that causes a shift in perspective and turns suicide into a more salient decision alternative (see Ward and Fox 1977). In this way the suicides of some can influence the decision-making processes and suicide risks of others.

Interaction Domains and Suicide • 717

Previous Research on Exposure Effects Previous research (Agerbo 2005; Qin, Agerbo and Mortensen 2002; Runeson and Åsberg 2003) clearly suggest that high suicide risks are associated with family histories of suicidal behavior. Less is known about the mechanisms that may contribute to this. In addition to the mechanisms discussed above, familial transmission of psychiatric disorder is likely to be important (Egeland and Sussex 1985). Results from twin and adoption studies furthermore suggest an important genetic component to suicide (Brent and Mann, 2005). Much higher levels of concordance have been reported between monozygotic than between dizygotic twins (e.g., Roy and Segal 2001), and Fu et al.’s (2002) large-scale twin study suggests that genetic factors account for as much as 36 percent of the variation in suicide ideation and 18 percent of the variation in suicide attempts. The remaining variation nevertheless is substantial, which suggests that nongenetic factors also are of considerable importance. Turning to the effect of exposure to suicides of non-family members, studies on the time-space clustering of suicides provide some indirect evidence on the importance of exposure effects. For example, Gould et al. (1990) found significant time-space suicide clusters among young people in the United States and estimated that such clusters accounted for about 5 percent of all teenage suicides. Another source of indirect evidence on exposure effects is the research on the effects of mass media reports on suicides. Phillips (1974) conducted the first large-scale study of mass media effects on suicide. He examined monthly statistics on suicides in the United States from 1947 through 1968 and found that the publication of front-page newspaper articles on suicides led to subsequent increases in the number of suicides (Phillips 1979,1980). Although some studies have failed to replicate these findings (e.g., Horton and Stack 1984; Kessler et al. 1989), results from a large number of studies have converged in support of the existence of a systematic association between media reports on suicides and subsequent suicide rates (Pirkis and Blood 2001; Stack 2000). In particular, suicide risks appear to be more influenced by reports of celebrity suicides than non-celebrity suicides: a recent meta-analysis of 419 studies showed that studies which looked at celebrity suicides were more than five times as likely to report an imitation effect than the others (Stack 2005). One possible reason for why copy-cat suicides are more likely after celebrity suicides is that such stories are more widely reported and therefore reach large audiences, but it may also reflect that people identify more strongly with well-known persons than with those they know nothing about.  Case studies of suicides in schools have similarly found that a suicide typically leads to elevated rates of suicide and suicide attempts (Brent et

718 • Social Forces 87(2)

al. 1989; Poijula, Wahlberg and Dyregrov 2001). Studies using survey data also suggest the existence of such effects (Bearman and Moody 2004; Cerel, Roberts and Nilsen 2005; Hazell and Lewin 1993; Ho et al. 2000; Lewinsohn, Rohde and Seeley 1994). Interaction Domains and Suicides The research summarized in previous sections thus suggests that a systematic causal relationship is likely to exist between the suicides of some and the suicide risks of others. Exposures, triggering events and the like are not randomly distributed in the population, however; in order to understand why some individuals are more likely than others to commit suicide, it appears essential to also consider the social structure in which the individuals are embedded. To clarify, consider the hypothetical ego-centered network in Figure 2. Person ℯ is directly tied to 10 other individuals (in reality the relevant ego networks are much larger than this, of course). Given the interaction mechanisms discussed above, if this person were to commit suicide, we would expect it to directly affect the suicide risks of the alters. The total effect of e’s suicide on the suicide rate, Ie, will depend on the number of alters who become aware of e’s suicide and the effect that e has on each of these alters. If we let pea represent the influence of e on each of the ne alters to which e is directly linked, the total direct influence of e is equal to5 ne

∑

I e = � peaea a =1

Person-to-person networks such as that in Figure 2 are particularly useful to focus on when patterns of interaction are fairly stable over time. But as suggested by Hedström (2005), this is not always the case. Individuals move between neighborhoods, schools, workplaces, etc., and during any given period of time, individuals typically interact only with a subset of the potential interaction partners in each domain. The amount of time devoted to interactions with people at work, for example, may be more or less stable over time, even though the specific people we interact with may vary a great deal. Similarly, the influence that people at work have on us may be more or less stable over extended periods of time, although the specific people who exercise this influence may vary a great deal. In cases like these, when the focus is on dynamic processes that unfold in networks that themselves are changing over time, it is often more appropriate and analytically more straightforward to focus on so-called catnets. Following White (1965:3), a catnet is a network describing the relations that exist between social categories, a category being a “bunch of people alike in

Interaction Domains and Suicide • 719

Figure 2. Hypothetical Ego-Centered Network Figure 2. Hypothetical Ego-centered Network

e

== Family == Neighbor == Co-worker

some respect.” Figure 3 contains an ego-centered domain-specific catnet representation of the person-to-person network of Figure 2. In this hypothetical example there are thus three relevant categories or interaction domains, and the influence of e’s suicide on the suicide rate, Ie, can be expressed as follows: K

∑

I e = � p ekek × � neekk k =1

where k indexes the domain, K the possible number of domains, peek k the average influence of individual e upon those in domain k, and n ek ek the number of individuals in domain k that are tied to e (excluding e). The value of Ie shows the extent to which the effect of a suicide is amplified by the social structure. If Ie is greater than zero we have a positive social multiplier, and the greater Ie is the more amplified the effect of e’s suicide will be. The extent of amplification depends upon the specific combination of peek k and nek values. Individuals in certain domains, Figure 3. Ego-Centered Catnet Representation of Figure 2 Figure 3. Ego-centered Catnet Representation of Figure 2 Domain 2: Work n2 = 5 pe2 Domain 1: Family n1 = 1

pe1

e

pe3

Domain 3: Neighborhood n3 = 4

720 • Social Forces 87(2)

the immediate family for example, are likely to be greatly affected by e’s suicide (the peek k value is likely to be high for this domain), but the number of individuals in this domain (nek) is likely to be rather low. In other domains the opposite pattern will hold; the peek k values will be low and the nek values will be high. Previous studies of exposure effects have typically considered only a single domain and, as a consequence, they have not been able to consider the variation and co-variation in the peek k and nek values. In the empirical analyses presented below, we focus on how these parameter values differ between the family domain and the work domain. From a public-health perspective it is the specific combination of values that matters. That is to say, even if the peek k -value is low, the public health effect can be considerable if the nek-value is high, implying that many individuals are exposed. And vice versa if the nek-value is low. The Stockholm Database The database we use in this study contains information on the entire adult population (age range 18-64) in the larger Stockholm metropolitan area for each year throughout the 1990s. Statistics Sweden assembled it by merging various administrative and population registers. The data is of high quality and missing values are virtually nonexistent. Constructing a database like this is only possible in a country where all government registers use the same identifying keys, in this case a personal ID number, and when government authorities regularly register a large amount of information about the individuals they come into contact with. The database includes a range of demographic and socio-economic information, information on both current and past family relations as well as several generations of kinship relations derived from parentchild information. It also includes information on places of work which will be used to link individuals to each other. The data used in this study consists of all individuals in the database who lived and worked in the Stockholm metropolitan area at some point during the years 1991-1999, and were in the age-range 18-64 at that point in time. In total, 1,195,098 individuals are included in the analyses, and each individual is observed over an average of 5.6 years. The outcome variable of interest refers to suicides and not to suicidal ideations. Information from the National Cause of Death Register was used to identify the suicide cases. We follow the usual practice in suicide epidemiology and define a suicide on the basis of the following cause-of-death codes: E950-E959 or E980-E989 for the years 1991-1996 (International Classification of Diseases, 9th revision) and X60-X84, Y87.0, or Y10-Y34 for the years 1997-1999 (International Classification of Diseases, 10th revision). Because we use lagged independent variables, the suicides

Interaction Domains and Suicide • 721

Figure 4a. Number of Individuals at Risk Figure 4a. Number of Individuals at Risk 500000 Number of Individuals at Risk

450000 400000 350000 300000

Women

250000

Men

200000 150000 100000 50000 0

92

93

94

95

96

97

98

99

Year

Figure Figure4b. 4b.Number NumberofofSuicides Suicides 140

Number of Suicides

120 100 80

Women Men

60 40 20 0

92

93

94

95

96 Year

97

98

99

722 • Social Forces 87(2)

that we seek to explain took place from 1992 through 1999. All in all 1,116 suicides took place in the study population during these years. Figure 4a and 4b shows how the number of suicides and the number of individuals at risk changed during the 1990s. About 400,000 men and 400,000 women were at risk each year and between 54 and 118 men and 46 and 59 women committed suicide each year. Methods Since the events to be analyzed are recorded on an annual basis only, a discrete-time event history approach is used. We estimate the parameters of the following type of model:

where pit equals the hazard rate, or the conditional probability that individual i will commit suicide during year t , given that s/he was alive at the beginning of year t, and α, γk, ,  and λ are logistic regression coefficients to be estimated. xikt are individual-level risk attributes likely to influence an individual’s suicide risk. Fit-1 is a variable measuring the number of suicides that took place in individual i’s family during the preceding year, and Wit-1 is a variable measuring the number of suicides that took place in the individual’s workplace during the preceding year. Family members were identified in two steps. First we identified the focal individual’s partner, mother, father, siblings, children, grandparents and grandchildren. Then we identified the mother, father, siblings, children, grandparents and grandchildren of the family members identified in the first step. All the individuals identified in these two steps were regarded as family members of the focal individual.6 “Workplace” is defined as the specific establishment in which the individual worked at the end of November during the preceding year. Before the parameters of the above model are estimated, the units must be changed from persons to “person years” so that each person contributes as many observations as the number of years that he was at risk (see Allison 1982). For example, an individual who entered our population in 1998 will only contribute one observation, while an individual who entered in 1994 will contribute five observations. The set of 1,195,098 persons included in these analyses contributed a total of 6.65 million person years. The set of right-hand variables included in the analysis was guided by findings from prior research. As background information to the eventhistory analyses, these variables are described in tables 1a and 1b.

Women 410

-1.78 -6.43 .00 -6.64 -6.82 -10.16

Men 706

-2.63 -6.45 -5.68 -.78 -3.86 -1.44

Actual Number of Suicides Removing Effect of: Family exposure Workplace exposure � 100 employees � 100 employees Widowhood Divorce

.22 .21 .00049 .00046 Number of exposures to suicide in family Exposure to suicide in family

.44 .11 .070 Number of exposures to suicide at workplace

.33

Women .053 Men .051 Women .0027 Men .0025 Definition Number of exposures to suicide at workplace in

Exposure Variables Exposure to suicide at workplace with less than 100 employees Exposure to suicide at workplace with more than 100 employees

Above we discussed in some detail the reasons for including the exposure variables. Let us also in a highly telescoped manner mention a few words about why the control variables described in Table 1b were included in the analysis. As far as age and gender are concerned, Girard (1993) and many others have shown that they are systematically related to suicide risks even when controlling for a host of other factors. The family type variables are similarly included because previous research has shown that living alone is associated with an elevated suicide risk (Heikkinen et al. 1995). Having young children has been found to be associated with lower suicide risks, especially among women (Qin, Agerbo and Mortensen 2003), but the protective effect appears to decrease with the age of the child (Cantor and Slater 1995). The country-of-birth variables are included because Johansson et al. (1997) found that foreign-born Swedes had higher suicide risks than the native born. Lorant et al. (2005) found that low educated men had higher suicide rates than highly educated men, and having a low income has similarly been found to be associated with a higher risk of suicide (Qin, Agerbo and Mortensen 2003). Permanent disability or sickness has similarly been found to be associated with an increased suicide risk (Lewis and Sloggett 1998), and so has sickness-related absence from Table 3: Estimates of the Public-Health Effects of Suicide Exposures and Selected Comparison Events Measured in Number of Suicides

Women 99.74 .25 .01 92.28 5.25 1.54 .79 .15 99.95 .05 .00 Men 99.75 .25 .00 94.64 4.14 .77 .41 .04 99.95 .05 .00

Distribution (%)

Number of Exposures 0 1 2 0 1 2 3 4 0 1 2 Standard Deviation Mean

Table 1a: Description of Exposure Variables Included in the Analysis Table 1a: Description of Exposure Variables Included in the Analysis

Table 3 had < 100 where it should have been � 100

Interaction Domains and Suicide • 723

Early retirement Social welfare recipient Divorce Widow(er) Unnatural death in family

Children aged 0-3 years Children aged 4-15 years Country of birth

Variable Name Family type

Definition Dummy variables distinguishing between: Couple living together One-person unita Other At least one child aged 0-3 living at home At least one child aged 4-15 living at home Dummy variables distinguishing between: Sweden Finland Denmark, Norway, Iceland Easter Europe other than former Yugoslavia Australia, New Zealand, Oceania Other Retired early due to a medical reason Received social welfare benefit Divorced less than 5 years ago Became widow(er) less than 5 years ago Exposed to unnatural death of family member due to injury, homicide, or war 50.3 18.22 31.46 13.6 28.55 84.55 6.31 .83 1.62 .03 6.67 3.1 5.3 3.6 .36 .04

50.4 11.9 37.7 13.3 25.0 84.55 4.14 .77 1.30 .00 9.24 1.9 5.2 3.5 .15 .03

Distribution (%) Men Women

Table 1b: Description of Control Variables Included in the Analysis Table 1b: Description of Control Variables Included in the Analysis 724 • Social Forces 87(2)

work (Qin, Agerbo and Mortensen 2003). Research on negative life events such as unnatural deaths and divorces similarly suggest that exposure to such events tends to elevate the suicide risk (Heikkinen, Aro and Lonnqvist 1994; Zisook, Chentsova-Dutton and Shuchter 1998), and this also appears to be the case with widowhood (Agerbo 2005; Luoma and Pearson 2002).7 Results

The logistic regression results are presented in tables 2a and 2b, and we report both the unadjusted bivariate associations between each covariate and the risk of committing suicide, and the adjusted ones which control for the effects of all the other covariates. We start by summarizing the effects of the exposure variables, and unless specifically noted, the discussion refers to the effects that have been adjusted for the effects of the other covariates.8

0-27 1-8453 0-29 1-8453 3.8 1392.0 8.7 8.8 3.8 Workplace size 396.8 558.5 986.9 Notes: a Notes: This category may include cohabiting couples who do not have any children in common. a b This category may include cohabiting couples who do not have any children in common. One base-amount is approximately equal to 33,000 SEK b One base-amount is approximately equal to 33,000 SEK

0-7.4 0-8.8 .46 1.1 1.2

.52

0-8.7 0-8.7 1.9 1.7 1.0 .73

Family size

Age Educational level Benefits for illness or work related injuries Income

Table 1b: Continued Table 1b continued

CA, 1392.0 is aligned now.

Age in November Years of education Logged amount (100’s SEK) due to illness or work related injuries received Logged total disposable income (in baseamountsb) in year t Number of family members in the extended family (not including the focal individual) Number of employees at the workplace

Men 39.7 10.0

Mean Women 39.8 10.0

Standard Deviation Men Women 12.2 12.4 5.4 5.4

Men 18-65 0-18

Range Women 18-65 0-18

Interaction Domains and Suicide • 725

We included family size and workplace size in the model to control for the possible effects of the size of the interaction domain. We had not expected to find an effect of family size, but we had anticipated the possibility of finding a positive workplace size effect because it seemed plausible that it could be the proportion rather than the absolute number of suicides that mattered. The results are the reverse, however. We find that individuals from large families run a higher risk of committing suicide than indivi­ duals from smaller families, and that the size of the workplace has no independent effect on the suicide risk after controlling for the other covariates in the model (and distinguishing between exposure in workplaces with more and with less than 100 employees). The exposure variables behave more or less as expected. This is particularly true for exposure within the family. The odds ratios are extremely high for women as well as for men, suggesting an eight- to nine-fold elevated risk for those who have been exposed to the suicide of a family member.9 It is important to keep in mind that this “social-interaction effect” is likely to be upwardly biased, however. The suicide propensities of two members of the same family are likely to be associated because of common ancestry as well as unmeasured common environmental influences. Such factors are not properly captured

Variable Number of Exposures to Suicide In family At workplace with � 100 employees At workplace with � 100 employees Age Years Years2/1000 Family Type Couple 1-person Other Children No Aged 0-3 yrs Aged 4-15 yrs Education Level Years of education Years of education2/1000 Early Retirement No Yes Benefits for Illness or Work Related Injuries (logged) Social Welfare Recipient 1.58*** .28***

1.00 4.87*** 1.33***

1.03 996.5*

3.75-6.34 1.29-1.36

.98-1.08 993.4-999.6

.30-.59 .37-.58

1.00 .42*** .46***

-.86*** -.77*** .03 -.04*10-1*

2.30-3.20 .96-1.68

1.01-1.11 999.0-1.00

2.71*** 1.27

1.06** 999.5

.06** -.00

3.00-28.25 2.06-9.31 .91-1.36

95%CI

.79*** .23***

-.04 2.69

-.25 -.35*

1.19*** .69***

.09*** -.80**

2.12*** 1.25*** .01

b

n = 601,245; N = 3,338,578

9.98*10-1*** .24

9.21** 4.38** 1.11

2.22*** 1.48*** .11

b

Unadjusted OR

Men

Table 2a: Parameter Estimates of Discrete Event-History Models. Men. Table 2a: Parameter Estimates of Discrete Event-History Models for Men

1.00 2.21*** 1.25***

.96 1002.7

1.62-3.02 1.21-1.30

.92-1.00 999.7-1005.7

.54-1.13 .54-.92

2.61-4.14 1.45-2.75

1.00 3.28*** 2.00*** 1.00 .78 .70*

1.04-1.15 998.6-999.8

2.70-25.64 1.62-7.56 .78-1.31

95%CI

1.09*** 999.2*

8.33*** 3.50*** 1.01

Adjusted OR

726 • Social Forces 87(2)

2.63-13.13 .62-31.36 1.19-2.17 1.00-3.50 .29-1.49 .69-1.19 .99-1.03 .99-1.00

1.00 5.87*** 1.00 4.41 1.00 1.60** 1.87* .67 .91 1.01 1.00

1.77*** 1.48

-.10 .01 .00

-

1.30-2.42

1.00 1.77***

.57***

.47** .63* -.40

.76-.97

.86*

-.16*

2.18-3.39

1.00 2.72***

9.97*10-3***

.32* .66* -.51 -.08 .07*** .00 -11.69*** .05 701.51

1.48

1.03*

.04

-.35***

.61***

1.00 1.37* 1.93* .60 .92 1.07*** 1.00

1.00 4.40

1.00 2.80*

1.00 1.04

.70***

1.00 1.85***

1.01-1.87 1.04-3.60 .27-1.35 .69-1.23 1.05-1.10 .99-1.00

.61-31.50

1.25-6.30

.74-1.44

.64-.78

1.45-2.35

Note: Estimated coefficients (b), odds ratios (OR) and confidence intervals for the odds ratios (CI). n = number of individuals and N = number of person years.

No Yes Level of Income Logged annual amount Divorcee No Yes Widow/Widower No Yes Unnatural Death in Family No Yes Country of Birth Sweden Finland Denmark/Norway/ Iceland Eastern Europe other than Yugoslavia Australia/New Zealand/Oceania Other Family Size Workplace Size Constant Pseudo R2 -2LL

Interaction Domains and Suicide • 727

Variable Number of Exposures to Suicide In family At workplace with � 100 employees At workplace with � 100 employees Age Years Years2/1000 Family Type Couple 1-person Other Children No Aged 0-3 yrs Aged 4-15 yrs Education Level Years of education Years of education2/1000 Early Retirement No Yes Benefits for Illness or Work Related Injuries (logged) Social Welfare Recipient 2.41-38.11 .14-6.09 1.08-1.51

95%CI

1.92*** .32***

1.00 6.84*** 1.38***

1.00 998.8

5.32-8.80 1.33-1.44

.94-1.07 995.7-1002.9

.19-.50 .37-.63

1.00 .31*** .48***

-1.18*** -.73*** .00 -1.22

2.68-4.21 2.68-4.21

1.00 3.36*** 1.73***

1.21*** .55***

1.11*** .28***

-.08** 4.95*

-.81** -.42*

1.22*** .50**

.08* -.92*

2.20** -.19 .10

b

n = 593,853; N = 3,308,271

.99-1.12 998.9-1.00

2.28*** -.08 .24** 1.06 999.6

9.79*** .92 1.28**

b

.05 -.42

Unadjusted OR

Women

Table 2b: Parameter Estimates of Discrete Event-History Models. Women. Table 2b: Parameter Estimates of Discrete Event-History Models for Women

1.00 3.05*** 1.32***

.92** 1005.0*

1.00 .45** .66*

1.00 3.40*** 1.65**

1.09* 999.1*

9.01** .83 1.11

.26-.76 .47-.91

2.55-4.53 1.18-2.31

1.02-1.16 998.3-999.9

2.23-36.30 .12-5.55 .86-1.42

95%CI

2.23-4.15 1.26-1.39

.87-.98 1000.9-1009.1

Adjusted OR

728 • Social Forces 87(2)

.80-1.18 1.60-3.29 3.69-12.95 3.43-55.25 1.49-2.72 .47-3.37 1.55-4.24 1.08-54.82 .32-.93 .99-1.03 1.00-1.00

.97 1.00 2.29*** 1.00 6.91*** 1.00 13.77*** 1.00 2.01*** 1.26 2.57*** 7.70* .54* 1.01 1.00**

-.03 .83*** 1.93*** 2.62*** .70*** 0.29 .94*** 2.04* -.61* .01 .09*10-4**

2.73-4.59

1.00 3.54***

1.26***

.69** .31 .89** 2.41* -.54 .08*** .00 -12.38*** .08 652.85

2.47***

1.15***

.38

-.29**

.96***

1.00 1.99** 1.37 2.44** 11.12* .59 1.08*** 1.00

1.00 11.80***

1.00 3.15***

1.00 1.47

.75**

1.00 2.60***

1.42-2.80 .51-3.67 1.44-4.11 1.62-76.28 .34-1.02 1.05-1.11 .99-1.00

2.89-48.17

1.64-6.05

.99-2.15

.61-.92

1.95-3.47

Note: Estimated coefficients (b), odds ratios (OR) and confidence intervals for the odds ratios (CI). n = number of individuals and N = Note: Estimated (b), odds ratios (OR), and confidence intervals for the odds ratios (CI). n = number of individuals and N = number of personcoefficients years. number of person years. *p < .05; **p <.01; ***p < .001

No Yes Level of Income Logged annual amount Divorcee No Yes Widow/Widower No Yes Unnatural Death in Family No Yes Country of Birth Sweden Finland Denmark/Norway/ Iceland Eastern Europe other than Yugoslavia Australia/New Zealand/Oceania Other Family Size Workplace Size Constant Pseudo R2 -2LL

Interaction Domains and Suicide • 729

730 • Social Forces 87(2)

by our control variables, and this is likely to inflate the estimate of the family-based social-interaction effect. Exposure at work is only significant for men, and only in workplaces with fewer than 100 employees. The effect of workplace exposure is not of the same magnitude as that of family exposure, but it is nevertheless considerable (OR = 3.50). The workplace exposure effect is unconfounded by genetic and other shared factors, and thus is likely to be caused by the type of social mechanisms discussed above. The reason for the absence of an effect in larger workplaces most likely does not mean that small and large workplaces differ from one another in this respect. More likely, it simply reflects the fact that in large workplaces the relevant reference group is not the workplace as a whole but some smaller department or subsection that we cannot identify in our data. As far as the reasons for the lack of a significant workplace effect among women, we can only speculate because the data does not allow for further analyses. Gender differences in workplace attachment appear to be one plausible reason. Although Swedish women are more likely to work than are women in many other countries, they still bear the main household and childcare responsibilities (Hall 1992), which may lead to a weaker attachment to the workplace, and work is typically more important for men’s self-identities than for women’s (Girard 1993; Neugarten and Hagestad 1976). As far as we know, this is the first study that systematically examines the effect of suicide exposures in the workplace. One possible objection to the way in which we have interpreted these results is that they may, at least in part, reflect unmeasured differences between workplaces in the exposure to other types of risk factors – stress levels, for example – which could possibly generate a spurious exposure effect. To partially control for that possibility we re-estimated the models in Table 2, including 58 industry dummies to partially control for unobserved heterogeneity. The results of these analyses were in every essential respect identical to those reported in Table 2 and they thus give further weight to the exposureinterpretation of the results reported here. If we then turn to the control variables, their effects are more or less as we would expect them to be. As previous research suggests, age is systematically associated with the suicide risk. The partial effect estimates suggest that the suicide risk gradually increases with age and that it is as highest at the age of 44 for men and 41 for women. As expected, coupled individuals, men as well as women, have lower suicide rates than singles. On the basis of these data we cannot tell whether this is a “protective effect” of the family arrangement as such or a selection effect due to suicidal individuals being less likely to be coupled. Also as expected, having children living at home appears to be associated with lower suicide risks for men as well as for women. For

Interaction Domains and Suicide • 731

men, the effect is significant only for children in the older age category, while for women we find the expected pattern of a declining “protective” effect with the age of the child. Contrary to much of previous research, we find that the net effect of years of education is positive for women (this can be seen by plotting the combined effect of the educational variables). That is, women’s suicide risk increases with increasing education, all else being equal. For men, the point estimates suggest the same pattern, but these estimates are not significantly different from zero.10 The variables that serve as proxies for health-related problems – early disability retirement and illness and work related injuries – behave as expected. Individuals who received such benefits had clearly elevated risks. The same is true for the two deprivation-related variables. The regression coefficient associated with the social-welfare variable is positive, and the regression coefficient of the income variable is negative. Having gone through a divorce in the past five years, somewhat surprisingly given previous research findings, does not appear to significantly elevate the risk for either sex. It should be emphasized that this refers to the partial effect of divorce, however. As the results of the bivariate unadjusted analyses show, those who had experienced a divorce had a significantly higher suicide risk than those who had not experienced a divorce. Hence, these results suggest that the two groups differ from one another in relevant respects, and after controlling for these differences, the divorce experience is no longer significantly related to the suicide risk. Having become a widow or widower in the past five years is associated with an elevated risk for men as well as for women, and this is also true for those who experienced an unnatural death in the family. This latter variable is associated with a sharply elevated risk for women. The point estimate is high also for men, but the test suggests that the estimate is not significantly different from zero. Finally, and as expected, immigrants from several different nations have elevated suicide risks as compared to those born in Sweden. With data like this it is not possible to assess to what extent this is a result of the immigration and assimilation experience as such, or whether it reflects individual differences between immigrants and native Swedes not picked up by the variables included the model. For our purpose, this is not a major concern, however, because we only included country of birth as a control variable to reduce the risk of the estimates of the exposure effects becoming confounded. To assess the overall importance of networks it is not sufficient to examine only social-interaction effects like those in Table 2, one must also take into account the number of individuals who are exposed. This is particularly important when comparing domains such as these that differ widely in size

732 • Social Forces 87(2)of the Public-Health Effects of Suicide Exposures and Selected Comp Table 3: Estimates

Number of Suicides Table 3: Estimates of the Public-Health Effects of Suicide Exposures and Selected Comparison Events Measured in Number of Suicides Actual Number of Suicides Removing Effect of: Family exposure Workplace exposure � 100 employees � 100 employees Widowhood Divorce Unnatural deaths in family

Men 706 -2.63 -6.45 -5.68 -.78 -3.86 -1.44 -.77

Women 410 -1.78 -6.43 .00 -6.64 -6.82 -10.16 -1.83

and average degree. On average, the suicides being studied here resulted in 2.9 family members being exposed, while suicides at small workplaces, here defined as workplaces with less than 100 employees, resulted in 15.3 other individuals being exposed.11 Table 3 displays results taking into account both the dyad-based interaction effects of Table 2 and the degree-based effect, i.e., the number of individuals being exposed to each suicide. The first row of Table 3 shows the actual number of suicides that took place, and this is identical to the number of suicides predicted by the logistic regression models. The second row shows what the number of suicides would have been according to the logistic models had no one been exposed to the suicide of a family member (or, equivalently, the number of suicides being observed if family-based exposure had no effect on the suicide risk). These estimates suggest that we would then have observed three fewer male suicides and two fewer female suicides than we actually observed. These public-health effects may seem surprisingly small given the considerable social-interaction effects shown in Table 2. Although one should treat these estimates with great caution and not pay too much attention to the exact magnitude of the estimates, they illustrate the importance of the distinctions introduced earlier. Although the social-interaction effects, what we referred to as peek k , are substantial, the collective public health effect is rather modest because the networks within the family domain are such that not many individuals are exposed; the nek values are low within the family domain. The third row of Table 3 performs the same kind of counterfactual thought experiment for workplace exposure. Since the effects were significant only for men, we focus exclusively on the public health effect among men. Had no one been exposed to a suicide in the workplace, these estimates suggest that there would have been six to seven fewer male suicides than actually was observed. A significant effect is found only in the smaller workplaces, but as argued above, this is most likely due to poor measurement of the

Interaction Domains and Suicide • 733

relevant reference groups in the larger workplaces. Had we been able to identify the relevant reference groups in the larger workplaces, undoubtedly the estimated public-health effect would have been considerably higher since more than 40 percent of employees work in such workplaces. For reasons such as these one should not pay much attention to the exact figures, but even with these caveats in mind, the results clearly suggest that it is important to consider the workplace when seeking to explain suicide rates. Although the peekk value for the work domain was much lower than for the family domain, the public-health effect of exposure in the work domain was much greater because of its higher nek value. The last three rows of Table 3 are included as comparison points for evaluating the substantive importance of the estimated public-health effects of the exposure variables. Here we perform the same type of thought experiments for three negative life events that are often considered in suicide research: becoming a widower, going through a divorce, and experiencing the unnatural death of a family member. For men, family exposure appears to be of approximately the same public-health importance as these three types of life events, while workplace exposure is more important. For women, the public-health effect of family exposure is comparable to that of an unnatural death of a family member, but it is not of the same magnitude as that of the other two types of events. Conclusion The analyses reported here suggest that network effects are important for explaining suicides. Suicides of family members as well as of coworkers influence the suicide risks of those exposed. In addition, the results suggest that the degree of the deceased individual’s egocentered network is important because it influences the number of exposed individuals and thereby the suicide rate. When examining the dyad-based social-interaction effects, exposure in the family domain seemed to be about twice as important as exposure in the work domain, but when we also took into account the degree-based exposure effects and examined the collective public-health effect, we found workplace exposure for men to be at least twice as important as family exposure. This result is particularly striking because the estimated family-exposure effect is likely to be upwardly biased since we were not able to control for factors such as the genetic component of suicide and inheritable psychiatric disorders known to increase the suicide risk. Individual-level studies of exposures to suicide have largely concentrated on adolescents, and it is generally believed that exposure effects are important only within this age group. As far as we know, this is the first study to demonstrate that workplace exposure is also associated with a

734 • Social Forces 87(2)

higher suicide risk among adults. National suicide prevention strategies usually pay no attention to suicide exposure in the workplace. The results reported here suggest that prevention strategies focusing on the workplace may be at least as important as those focusing on the family. We believe that some of the empirical results and analytical distinctions introduced in this article are important also for explaining other outcomes. Whenever the interaction among individuals is deemed important for the collective outcome to be explained, as in differential-association theory’s (Sutherland 1947) explanation of crime rates for example, the distinction between dyad-based and degree-based network effects appears essential. Focusing exclusively on one of these two types of effects can lead to a faulty understanding of the importance of social interactions and networks. In addition, we believe that this article illustrates the importance of tightly linking theory and data.12 If our theory suggests that networks and/ or localized social interactions are important, it is essential to use datasets which actually contain such information because the wider the gap is between theory and data the less theoretical bearing and relevance will the empirical research have. It is always tempting to cut corners by using aggregate data or indirect proxy variables from existing surveys, but since social action and interaction is at the heart of so much of sociological theory, there will often be no good alternative to collecting the type of data used in this article, i.e., longitudinal data on the actions and attributes of large numbers of individuals and the patterns of interaction among them.13 Notes 1. See Sainsbury, Jenkins and Levey (1980) and Breault (1986) for examples. Mäkinen’s (1997) replication study of Sainsbury et al.’s study, however, has cast serious doubts on the reliability and robustness of such aggregate associations. Also see Pescosolido (1994) for a critique of the approach. 2. Let us emphasize that analyzing suicides as intentional actions does not mean that the intention behind all suicides is to take one’s life. The intention may have been to send a signal to others about a desperate life situation, but the unintended outcome of the act was death. Nor does an action approach imply that individual intentions are necessarily unclouded by strong emotions or cognitive biases. 3. It also is exemplified by the title of the Swedish singer Magnus Uggla’s 1989 hit album “What’s the Point of Killing Yourself if You Can’t be Around to Hear the Yap.” 4. We ignore pure opportunity-related interaction mechanisms because a person’s suicide rarely affects others’ opportunities to commit suicide. One possible exception is when an individual invents a new suicide method, e.g.,

Interaction Domains and Suicide • 735

the use of car exhaust gas as a means of committing suicide in Britain in the 1970s. See Clarke and Lester (1987). 5. pea can be thought of as measuring how the probability of an alter’s committing suicide is affected by e’s suicide, holding everything else constant. 6. It should be observed that only individuals residing in the Stockholm metropolitan area are included in the database. 7. Given the fact that previous research suggests that religion plays an important role in this context, we regret the fact that the database includes no information on the individuals’ religious orientations. 8. It would have been useful to estimate fixed-effect logit models, but unfortunately such models would be highly inefficient because they disregard so much of the data. Suicides are extremely rare events, and fixed-effect logit models only make use of data on units with variation on the dependent variable. Since suicides occur so rarely, we also estimated the so-called rare-event logit model (King and Zeng, 1999). Qualitatively the results were identical to those reported here. 9. These ORs of family exposure are higher than those found in another Swedish register-based study (Runeson and Åsberg, 2003). There are some differences in the study designs that may explain these discrepancies, most notably that they used people who had died from other causes than suicide as their control group. 10. If we examine the relationship between the two educational variables and the suicide risks without any other control variables, the expected negative pattern is found. 11. In larger workplaces, the number of potentially exposed individuals was of course much higher, but because we do not have information on the way in which these workplaces were internally organized, we cannot estimate the likely number of truly exposed individuals in large workplaces. Our cut-off value of 100 to distinguish between “large” and “small” workplaces is arbitrarily chosen, but the results seem robust at least to small changes in this value. 12. See Hedström and Bearman (forthcoming) for more detailed discussions of this analytically oriented middle-range approach that seeks to tightly link sociological theory and empirical research. 13. Needless to say, data of the size and scale used in this article will be prohibitively expensive to collect, but such data only is needed when the outcomes to be explained are extremely rare, and even in such circumstances smaller case-control designs often can be used.

736 • Social Forces 87(2)

References Åberg, Yvonne. 2003. Social Interactions: Studies of Contextual Effects and Endogenous Processes. Stockholm: Department of Sociology, Stockholm University. Agerbo, Esben. 2005. “Midlife Suicide Risk, Partner’s Psychiatric Illness, Spouse and Child Bereavement by Suicide or Other Modes of Death: A Gender Specific Study.” Journal of Epidemiology and Community Health 59(5):407-12. Allison, Paul D. 1982. “Discrete-Time Methods for the Analysis of Event Histories.” Sociological Methodology 13:61-98. Bandura, Albert. 1994. “Self-Efficacy.” Pp. 71-81. Encyclopedia of Human Behavior. Vilayanur S. Ramachaudran, editor. Academic Press. Baron, James N., and Peter C. Reiss. 1985. “Same Time, Next Year: Aggregate Analysis of the Mass Media and Violent Behavior.” American Sociological Review 50(3):347-63. Bearman, Peter. 1991. “The Social Structure of Suicide.” Sociological Forum 6(3):501-24. Bearman, Peter, and James Moody. 2004. “Suicide and Friendships among American Adolescents.” American Journal of Public Health 94(1):89-95. Breault, Kevin D. 1986. “Suicide in America: A Test of Durkheim’s Theory of Religious and Family Integration, 1933-1980.” American Journal of Sociology 92(3):628-56. Brent, David A., Mary M. Kerr, Charles Goldstein, James Bozigar, Mary Wartella and Marjorie J. Allan. 1989. “An Outbreak of Suicide and Suicidal Behavior in a High School.” Journal of the American Academy of Child & Adolescent Psychiatry 28(6):918-24. Brent, David A., and J. John Mann. 2005. “Family Genetic Studies, Suicide and Suicidal Behavior.” American Journal of Medical Genetics Part C 133C(1):13-24. Cantor, Christopher H., and Penelope J. Slater. 1995. “Marital Breakdown, Parenthood, and Suicide.” Journal of Family Studies 1(2):91-102. CDC. 1994. “Centers for Disease Control Recommendations for a Community Plan for the Prevention and Containment of Suicide Clusters.” Morbidity and Mortality Weekly Report 37(S6):1-12. Cerel, Julie, Timothy A. Roberts and Wendy J. Nilsen. 2005. “Peer Suicidal Behavior and Adolescent Risk Behavior.” Journal of Nervous and Mental Disease 193(4):237-43. Clarke, Ronald V., and David Lester. 1987. “Toxicity of Car Exhausts and Opportunity for Suicide: Comparison between Britain and the United States.” Journal of Epidemiology & Community Health 41(2):114-20.

Interaction Domains and Suicide • 737

Cutler, David M., Edward L. Glaeser and Karen Norberg. 2001. “Explaining the Rise in Youth Suicide.” Pp.219-70. Risky Behavior among Youths: An Economic Analysis. Jonathan Gruber, editor. University of Chicago Press. Davidson, Donald. 1980. Essays on Actions and Events. Clarendon Press. Egeland, Janice A., and James N. Sussex. 1985. “Suicide and Family Loading for Affective Disorders.” Journal of American Medical Association 254(7):915-8. Elster, Jon. 1983. Explaining Technical Change: A Case Study in the Philosophy of Science. Cambridge University Press. Farberow, Norman L. 1975. “Cultural History of Suicide.” Pp. 1-15. Suicide in Different Cultures. Norman L. Farberow, editor. University Park Press. Fu, Qiang, Andrew C. Heath, Kathleen K. Bucholz, Elliot C. Nelson, Anne L. Glowinski, J. Goldberg, M.J. Lyons, M.T. Tsuang, Theodore Jacob, M.R. True and S.A. Eisen. 2002. “A Twin Study of Genetic and Environmental Influences on Suicidality in Men.” Psychological Medicine 32(1):11-24. Girard, Chris. 1993. “Age, Gender, and Suicide: A Cross-National Analysis.” American Sociological Review 58(4):553-74. Gould, Madelyn S., Sylvan Wallenstein, Marjorie H. Kleinman, Patrick O’Carroll and James Mercy. 1990. “Suicide Clusters: An Examination of Age-Specific Effects.” American Journal of Public Health 80(2):211-2. Hall, Ellen M. 1992. “Double Exposure: The Combined Impact of the Home and Work Environments on Psychosomatic Strain in Swedish Women and Men.” International Journal of Health Services 22(2):239-60. Hankoff, Leon D. 1961. “An Epidemic of Attempted Suicide.” Comprehensive Psychiatry 2(5):294-8. Hazell, Philip, and Terry Lewin. 1993. “Friends of Adolescent Suicide Attempters and Completers.” Journal of the American Academy of Child and Adolescent Psychiatry 32(1):76-81. Hedström, Peter 2005. Dissecting the Social: On the Principles of Analytical Sociology. Cambridge University Press. Hedström, Peter, and Peter Bearman. Editors. Forthcoming. The Oxford Handbook of Analytical Sociology. Oxford University Press. Hedström, Peter, and Richard Swedberg. Editors. 1998. Social Mechanisms: An Analytical Approach to Social Theory. Cambridge University Press. Heikkinen, Martti E., Hillevi M. Aro and Jouko K. Lonnqvist. 1994. “Recent Life Events, Social Support and Suicide.” Acta Psychiatrica Scandinavica Supplementum 89(s377):65-72.

738 • Social Forces 87(2)

Hezel, Francis X. 1985. “Trukese Suicide.” Pp. 112-24. Culture, Youth and Suicide in the Pacific: Paper from an East-West Center Conference. Francis X. Hezel, Donald H. Rubinstein and Geoffrey M. White, editors. University of Hawaii Press. Ho, Ting-Pong, Patrick Wing-leung Leung, Se-Fong Hung, Chi-Chiu Lee and ChunPan Tang. 2000. “The Mental Health of the Peers of Suicide Completers and Attempters.” Journal of Child Psychology and Psychiatry 41(3):301-8. Horton, Hayward, and Steven Stack. 1984. “The Effect of Television on National Suicide Rates.” Journal of Social Psychology 123(1):141-2. Jacobs, Douglas G., Margaret Brewer and Marci Klein-Benheim. 1999. “Suicide Assessment: An Overview and Recommended Protocol.” Pp. 3-39. The Harvard Medical School Guide to Suicide Assessment and Intervention. D.G. Jacobs, editor. Jossey-Bass. Johansson, Leena M., Jan Sundquist, Sven-Erik Johansson, Jan Qvist and B. Bergman. 1997. “The Influence of Ethnicity and Social and Demographic Factors on Swedish Suicide Rates. A Four-Year Follow-Up Study.” Social Psychiatry and Psychiatric Epidemiology 32(3):165-70. Kessler, Ronald C., Geraldine Downey, Horst Stipp and J. Ronald Milavsky. 1989. “Network Television News Stories about Suicide and Short-Term Changes in Total U.S. Suicides.” Journal of Nervous and Mental Disease 177(9):551-55. King, Gary, and Langche Zeng. 2000. “Logistic Regression in Rare Events Data.” Political Analysis 9(2):137-63. Lewinsohn, Peter M., Paul Rohde and John R. Seeley. 1994. “Psychosocial Risk Factors for Future Adolescent Suicide Attempts.” Journal of Consulting and Clinical Psychology 62(2):297-305. Lewis, Glyn, and Andy Sloggett. 1998. “Suicide, Deprivation, and Unemployment: Record Linkage Study.” British Medical Journal 317(7168):1283-6. Linehan, Marsha M., Judith L. Goodstein, Stevan L. Nielsen and John A. Chiles. 1983. “Reasons for Staying Alive When You Are Thinking of Killing Yourself: The Reasons for Living Inventory.” Journal of Consulting and Clinical Psychology 51(2):276-86. Luoma, Jason B., and Jane L. Pearson. 2002. “Suicide and Marital Status in the United States, 1991-1996: Is Widowhood a Risk Factor?” American Journal of Public Health 92(9):1518-22. McKenzie, Nigel, Sabine Landau, Nanveet Kapur, Janet Meehan, Jo Robinson, Harriet Bickley, Rebecca Parsons and Louis Appleby. 2005. “Clustering of Suicides among People with Mental Illness.” British Journal of Psychiatry 187(5):476-80.

Interaction Domains and Suicide • 739

Michel, Konrad, and Ladislav Valach. 1997. “Suicide as Goal-Directed Action.” Archives of Suicide Research 3(3):213-21. Mäkinen, Ilkka H. 1997. “Are There Social Correlates to Suicide?” Social Science and Medicine 44(12):1919-29. Neugarten, Bernice L., and Gunhild O. Hagestad. 1976. “Age and the Life Course.” Pp. 35-55. Handbook of Aging and the Social Sciences. Robert H. Binstock and Ethel Shanas, editors. Van Nostrand Reinhold Co. Pescosolido, Bernice A. 1994. “Bringing Durkheim into the Twenty-First Century: A Network Approach to Unresolved Issues in the Sociology of Suicide.” Pp. 264-96. Emile Durkheim: Le suicide, One Hundred Years Later. David Lester, editor. Charles Press. Phillips, David P. 1974. “The Influence of Suggestion on Suicide: Substantive and Theoretical Implications of the Werther Effect.” American Sociological Review 39(3):340-54. . 1979. “Suicide, Motor Vehicle Fatalities, and the Mass Media: Evidence toward a Theory of Suggestion.” American Journal of Sociology 84(5):1150-74. . 1980. “Airplane Accidents, Murder, and the Mass Media: towards a Theory of Imitation and Suggestion.” Social Forces 58(4):1001-24. Pirkis, Jane, and R. Warwick Blood. 2001. “Suicide and the Media. Part I: Reportage in Nonfictional Media.” Crisis 22(4):146-54. Poijula, Soili, Karl-Erik Wahlberg and Atle Dyregrov. 2001. “Adolescent Suicide and Suicide Contagion in Three Secondary Schools.” International Journal of Emergency Mental Health 3(3):163-8. Qin, Ping, Espen Agerbo and Preben B. Mortensen. 2002. “Suicide Risk in Relation to Family History of Completed Suicide and Psychiatric Disorders: A Nested CaseControl Study Based on Longitudinal Registers.” Lancet 360(9340):1126-30. . 2003. “Suicide Risk in Relation to Socioeconomic, Demographic, Psychiatric, and Familial Factors: A National Register-Based Study of All Suicides in Denmark, 1981-1997.” American Journal of Psychiatry 160(4):765-72. Roy, Alec, David Nielsen, Gunnar Rylander and Marco Sarchiapone. 2000. “The Genetics of Suicidal Behaviour.” Pp. 210-21. The International Handbook of Suicide and Attempted Suicide. K. Hawton and K. Van Heeringen, editors. John Wiley & Sons. Roy, Alec, and Nancy L. Segal. 2001. “Suicidal Behavior in Twins: A Replication.” Journal of Affective Disorders 66(1):71-74. Rudd, M. David 2000. “The Suicidal Mode: A Cognitive-Behavioral Model of Suicidality.” Suicide and Life Threatening Behavior 30(1):18-33.

740 • Social Forces 87(2)

Runeson, Bo, and Marie Åsberg. 2003. “Family History of Suicide among Suicide Victims.” American Journal of Psychiatry 160(8):1525-26. Sainsbury, Peter, J. Jenkins and A. Levey. 1980. “The Social Correlates of Suicide in Europe.” Pp. 38-53. The Suicide Syndrome. Richard Farmer and Steven Hirsch, editors. Croom Helm. Shneidman, Edwin S. 1969. On the Nature of Suicide. Jossey-Bass. Stack, Steven 2000. “Suicide: A 15-Year Review of the Sociological Literature Part II: Modernization and Social Integration Perspectives.” Suicide and Life Threatening Behavior 30(2):163-76. . 2005. “Suicide and the Media: A Quantitative Review of Studies Based on Nonfictional Stories.” Suicide and Life Threatening Behavior 35(2):121-33. Sutherland, Edwin. 1947. Principles of Criminology. 4th edition. J.B. Lippincott. Ward, J.A., and Joseph Fox. 1977. “A Suicide Epidemic on an Indian Reserve.” Canadian Psychiatric Association Journal 22(8):423-26. White, Harrison C. 1965. “Notes on the Constituents of Social Structure.” Unpublished manuscript. Department of Sociology. Harvard University. WHO. 2000. “Preventing Suicide: A Resource for Media Professionals.” Geneva: World Health Organization. Zisook, Sidney, Yulia Chentsova-Dutton and Stephen R. Shuchter. 1998. “PTSD Following Bereavement.” Annals of Clinical Psychiatry 10(4):157-63.