A multilevel study of neighborhood networks and suicide* Ka-yuet Liu Department of Sociology University of Oxford

Word Count: 7,454 *This study is part of a research project led by Peter Hedström, and I am extremely grateful for his support and guidance. I thank Peter Bearman, Marissa King, Ray Fitzpatrick, Christofer Edling, Monica Nordvik, Esben Agerbo, Pin Qin and Ann Stewart for useful comments. An earlier version of this paper was presented to the American Sociological Association Annual Meeting, Boston, 2008; I thank the session attendants for their thoughtful comments. The author is supported by the Swire Education Trust. Support from Sun Chan is gratefully acknowledged. Direct all correspondence to Ka-yuet Liu, Paul F. Lazarsfeld Center for the Social Sciences, Columbia University, 420 W 118th St, 8FL, Mail Code 3355, New York, New York 10027, U.S.A. Email: [email protected]

NEIGHBORHOOD NETWORKS AND SUICIDE 2

NEIGHBORHOOD NETWORKS AND SUICIDE: A MULTILEVEL STUDY

Abstract Results on the social correlates of suicide rates have been inconsistent. It is unclear whether neighborhoods have any effect on suicide over and above the characteristics of the individuals who live in them. The macro-micro links need to be clarified in order to explain the effects of neighborhoods on suicide rates and make sense of the inconsistent results. This paper tests the macro-micro mechanisms implicit in Durkheim’s work and expanded by recent network theories on suicide. The mechanisms imply that a high prevalence in the neighborhood should cause a protective factor to become more protective and a risk factor less detrimental. Such contextual interactions can even reverse the relationship between an attribute and suicide when neighborhood composition changes. Specific predictions are derived to test the proposed mechanisms with multilevel modeling. The data are from a unique dataset of all of the 1.4 million adults who ever lived in the 500+ meaningfully defined neighborhoods in Stockholm during the 1990s. With one exception, the ten cross-level interactions tested are of the predicted direction, and the effects concerning income and ethnicity are substantial. The results support that neighborhoods do matter to suicide risk, and such contextual interactions contribute to the mixed results from aggregate-level studies.

NEIGHBORHOOD NETWORKS AND SUICIDE 3

BIOGRAPHY Ka-yuet Liu is a Post-Doctoral Researcher at the Paul F. Lazarsfeld Center for the Social Sciences, Columbia University. Her doctoral research at the University of Oxford studies a range of social mechanisms that generate different patterns of suicide rates. Her broader research interests include: social interactions, social networks, micro-macro links, and non-contagious diseases. Currently she works with Peter Bearman’s group to study the social determinants of autism.

NEIGHBORHOOD NETWORKS AND SUICIDE 4

I.

INTRODUCTION Geographical variations in suicide rates have long intrigued sociologists. Since Durkheim’s

(1897[1951]) seminal work, numerous ecological studies have been conducted, yet the results are often inconsistent: Conflicting findings on the effects of neighborhood’s socioeconomic status on suicide rates have been reported (Rehkopf and Buka 2006). Breault (1994) also cautions that the ecological correlates of suicide in the U.S. are likely to be artificial relationships. Mäkinen (1997) questions whether social correlates of suicide truly exist when he fails to replicate the results from the well-known study by Sainsbury, Jenkins and Levey (1980). Given the problem of ecological fallacy (Robinson 1950) and the conflicting findings, it is uncertain whether area of residence has any causal effect on suicide rates over and above the characteristics of the people who live there. Clarifying the macro-micro mechanisms should allow us to better understand the effects of neighborhood contexts on suicide and to explain the inconsistent results. While macro-level relationships may be due to self-selection into neighborhoods, area of residence can conceivably have a causal impact on suicide through (1) shared environment (e.g., access to suicide means and health-care facilities), and/or (2) the social interactions among residents. The latter processes are of particular interest to sociologists. Manski (1993; 2000) distinguishes between two types of social interactions, namely endogenous and contextual. Endogenous interactions take place when the behavior of the agent varies with the level of the behavior in the population. Contextual interactions refer to situations in which an agent’s behavior varies with the exogenous characteristics of the group members (Manski 2000). Examples of exogenous characteristics are socioeconomic and ethnic compositions. Contextual interactions are the focus of this paper1 which examines how such

1

We have studied the effect of endogenous interactions on suicide using the same dataset. See Hedström, Liu & Nordvik (Forthcoming)

NEIGHBORHOOD NETWORKS AND SUICIDE 5

interactions generate, at least in part, the neighborhood patterns of suicide observed in Stockholm during the 1990s. Analyzing contextual interactions taps directly into the question of how social contexts can affect an individual’s propensity to commit suicide. Despite renewed interest in neighborhood effects in the study of problem behaviors and health-related outcomes (for review, see Sampson, Morenoff, and Gannon-Rowley 2002), few studies have looked at the effects of interactions between individuals and their social contexts on suicide risk. Whether such interactions are important is in dispute: while Neeleman and Wessely (1999) find significant interactions between neighborhoods’ ethnic compositions and the effects of individual ethnicity, three other studies yield no or limited support for interaction effects (Agerbo, Sterne, and Gunnell 2007; Martikainen, Mäki, and Blomgren 2004a; van Tubergen, te Grotenhuis, and Ultee 2005). The scarcity of suitable data is one reason for the lack of research. Only with large-scale, multilevel data can such interactions between individual- and neighborhood- characteristics be examined. The rarity of suicide deaths requires a large number of observations in numerous geographical units to yield sufficient statistical power. This study draws upon a database compiled by Statistics Sweden by merging administrative and population registers. It provides comprehensive information on the 1.4 million adults who ever lived in the Stockholm metropolitan area from 1991 to 1999. The data are generally of very high quality; non-responses and missing items are virtually non-existent. One common methodological problem faced by neighborhood studies is the lack of meaningfully defined neighborhoods, since only information on large administrative boundaries is typically available (Diez Roux 2001). Because social interaction mechanisms are unlikely to span large and heterogeneous areas, using large geographical units increases the risk of grouping highly heterogeneous areas together and may give the erroneous impression that such interaction processes

NEIGHBORHOOD NETWORKS AND SUICIDE 6

are of negligible importance. The Stockholm database contains a residential variable specifically created to identify small homogenous residential areas. This provides the meaningful geographical units within which social interaction processes are plausible. Moreover, the database contains information on all individuals living in Stockholm that allows for the construction of the neighborhood-level measures that are unavailable in standard government statistics. This studies tests the interaction mechanisms derived from Durkheim’s theory (1897[1951]), Gibbs and Martin’s status integration theory (1964) and Pescosolido and Geogianna’s (1989) network theory. In essence, an extension of such classic mechanisms predicts that the effect of an individual attribute on suicide risk should depend on the neighborhood context, i.e. whether the particular attribute is common or rare. This gives rise to a specific prediction testable with multilevel analysis: regardless of whether the attribute is a risk or a protective factor at the individual level, its interaction with the analogous neighborhood-level variable should always be negative. It also means that an attribute can change from a risk to a protective factor (or vice versa) when it becomes more common. The mechanisms are discussed in detail in the following section. Section III describes the data and variables used. Section IV explains the statistical methods. Section V provides the results; Section VI discusses the implications of the findings.

II.

NEIGHBORHOOD INTERACTIONS Interaction mechanisms that are crucial to suicide risk have long been implicit in the classic

sociological theories of suicide. According to Durkheim (1897[1951]), an individual’s relationship with society can be conceptualized in two dimensions, namely social integration and social regulation. High suicide rates are the results of too high/too low social integration or social

NEIGHBORHOOD NETWORKS AND SUICIDE 7

regulation. Thus, he distinguished between four types: altruistic suicides that occur when social integration is too high; egoistic suicides that are the results of too low integration; fatalistic suicides that occur when social regulation is too strong; and anomic suicides that are common when social regulation is too weak. His conjecture implies that individuals who live in communities at extreme levels of social integration and regulation will present a higher risk of suicide. Thorlindsson and Bjarnasson (1998) argue that Durkheim’s theory “presupposes a social psychological theory that he never explicitly develops” (p.94). Durkheim’s insistence in framing his theory at the societal level leaves open the question of how social integration and regulation affect individual suicide risk. Less developed are the details of his “relational” hypotheses (Moksony 1994; van Tubergen, te Grotenhuis, and Ultee 2005): how group characteristics affect individual suicidal risk differentially. Recent research has begun to examine the interplay between group and individual characteristics (Ellison, Burr, and McCall 1997; Pescosolido 1990; Pescosolido 1994; Pescosolido and Georgianna 1989). Pescosolido’s network theory (1990; 1994; 1989) reformulates Durkheim’s theory into a macro-micro approach. She proposes that the integrative power of religion arises from the power of a religious affiliation to engage individuals in a stable social network. Borrowing terminology from network theory, Pescosolido differentiates the effects of (1) network density, (2) strength of the social ties and (3) functions of the network. The effects of network characteristics on suicidal behavior are demonstrated by the work of Bearman and Moody (2004). Using unique network data, they show that girls who are socially isolated or associate with friends who are not friends with one another have a higher risk of suicidal ideation than girls who are embedded in cohesive social networks. Under the network analysis’s framework, a social network can have more than one function. This clarifies the confusions caused by Durkheim’s separate treatments of social regulation and

NEIGHBORHOOD NETWORKS AND SUICIDE 8

integration(Pescosolido and Georgianna 1989). The potential regulative function of networks is considered by Van Turbergen, te Grotenhuis and Ultee (2005) as an alternative explanation of the religion-suicide. Hence, network theories of suicide are largely grounded on the psychosocial mechanisms of social support and/or normative pressures. The protective effect of social support is confirmed by a large number of psychosocial studies (e.g., Heikkinen, Aro, and Lönnqvist 1994; Liu, Chen, Chan, Lee, Law, Conwell, and Yip 2006). Similarly, normative pressure against suicide has been consistently shown to reduce suicide risk (Linehan, Goodstein, Nielsen, and Chiles 1983; Malone, Oquendo, Haas, Ellis, Li, and Mann 2000). Since social networks are promoted by homophily (McPherson, Smith-Lovin, and Cook 2001), we should expect to find that social support and normative pressure against suicide have the most significant effects within homogeneous groups. Apart from social support and normative pressure, dissonance-related stress is another likely mechanism. Gibbs and Martin’s (1964) status integration theory suggests that role conflicts tend to be greatest between statuses (e.g., occupational and marital statuses) that are seldom occupied at the same time. In other words, individuals who deviate too much from the “average” of the “typical” are believed to have a higher than average suicide risk. Such dissonance-related stress can be generated by perceived differences in ethnicity and socioeconomic status. Dissonance is also suggested as the underlying mechanism for the negative effect of unbalance network structure on suicidal ideation (Bearman and Moody 2004). This interaction mechanism implies that if more neighbors are in a similar condition, the level of stress and the associated suicide risk are likely to be lower.2 Evidence from research on non-fatal suicide attempts is in line with the dissonance-related stress mechanism: Hawton et al. (2001) finds that people who live in wealthy areas are more likely to 2

Similar arguments have been raised in the literature on relative income inequality and health. It has been suggested that the perception of relative deprivation is responsible for the well-known relationship between relative income inequality and health (Wilkinson, 1966). But see Rodgers (2002) for a critique.

NEIGHBORHOOD NETWORKS AND SUICIDE 9

cite financial difficulties as a reason for deliberate self-harm. Platt and Kreitman (1990) show that the risk of self-harm among the unemployed in Edinburgh decreased when the unemployment rates rose in the 1990s. Combining the effects of social support, normative pressure and dissonance, a plausible relational mechanism of suicide appears to include the following components: (1) the extent of social support and normative pressure is stronger among homogenous than heterogeneous groups, whereas dissonance-related stress is weaker among homogenous than heterogeneous groups; (2) this implies that individuals with many similar neighbors have better social support, stronger normative pressures and less dissonance-related stress; (3) in turn, this is likely to reduce the risk of suicide; (4) there is likely to be a decreasing rate of return such that when an attribute becomes common, the added benefits of having yet another neighbor similar to oneself will diminish. These interactions between individual and neighborhood characteristics are likely to be especially pronounced when the attribute of concern is relatively rare in the neighborhood. This simple extension of Durkheimian theories points to some specific predictions, as Figure 1 illustrates. Let X be an attribute that can affect suicide risk (e.g., an ethnic minority status). Because the aforementioned social mechanisms operate independently from the effect of the attribute on suicide, they should apply whether the attribute is a risk or protective factor. In Figure 1a, X is a risk factor (i.e. a person with X always has a higher suicide risk than someone without X). In Figure 1b, X is a protective factor (i.e. a person with X always has a lower suicide risk than someone without X). In both cases, if a person with X is more apt to interact with other members with the same attribute, the three aforementioned mechanisms will reduce the suicide risk of this individual when the proportion of neighbors with X increases. Simply stated, the presence of more similar others reduces the risk associated with X in Figure 1a, while it strengthens the protective

NEIGHBORHOOD NETWORKS AND SUICIDE 10

effect in Figure 1b. This leads to a negative relationship between the suicide risk associated with X and the proportion of neighbors with X in both scenarios. This is represented by the negative slopes for those with X in both Figures 1a and 1b. An increase in the proportion with X in the neighborhood may also cause the suicide risk among those without this attribute to increase, although the change may be indiscernible depending on the neighborhood’s actual composition. This is represented by the slightly positive slopes for a person without X in Figures 1a and 1b. With such interactions between individual and neighborhood characteristics, we can expect that X may even change from a risk to a protective factor (or vice versa) as the neighborhood’s composition changes (Figure 1c). ------------------------------------Figure 1 and Figure 2 Here ------------------------------------Such relationships at the individual-level indicate interesting relationships between the average suicide risk in the neighborhood and neighborhood compositions. Given the functional forms at the individual-levels and the relative sizes of the groups, Figure 2 depicts the relationships between the neighborhood-specific suicide risk and neighborhood composition under the different scenarios. In the first scenario (Figure 1a) in which X is generally a risk factor, an increasing proportion of residents with X gradually drives up the average suicide risk at the neighborhoodlevel. The negative cross-level interaction, however, dampens the rise and produces the curvilinear pattern shown in Figure 2a. The reverse occurs when X is generally a protective factor (Figures 1b and 2b). These two processes can combine to produce a U-shaped relationship (Figure 2c). This pattern coincides with the observations of

Neeleman and Wessely in South London (1999):

comparing the suicide risk among ethnic minorities and Caucasians, the risk among ethnic minorities

NEIGHBORHOOD NETWORKS AND SUICIDE 11

is lower in the election wards with large minority populations and higher in the wards with smaller minority populations. Figure 2 clearly illustrates that the relationships between suicide rates and neighborhoodlevel correlates may depend on the range of neighborhoods in the sample, which may explain the inconsistent results concerning the neighborhood correlates of suicide rates. As mentioned above, the additional benefits of having similar neighbors may begin to saturate when there are many neighbors belonging to the same group. It should be noted that the diminishing benefits, as Figure 3 illustrates, can only reduce the rate of change in the slopes depicted in Figures 3a-c, but would not alter the overall patterns. ------------------------------------Figure 3 Here -------------------------------------

Testing the Contextual Mechanisms with Multilevel Analysis The three mechanisms described above predict that regardless of whether the attribute is a risk or a projective factor at the individual level its interaction with the analogous neighborhood-level variable should always be negative. When an attribute increases suicide risk (a positive slope at the individual level), a negative interaction with its prevalence at the neighborhood would dampen this slope when the proportion of individuals with the attribute increases. When an attribute is protective (a negative slope at the individual level), a negative interaction with its prevalence would make its effect more negative at higher levels of prevalence. In multilevel analysis, this is equivalent to predicting that the cross-level interaction between an attribute and its prevalence in the

NEIGHBORHOOD NETWORKS AND SUICIDE 12

neighborhood should always be negative. Moreover, an attribute can change from a risk to a protective factor when it becomes more common in the neighborhood.

III.

DATA AND VARIABLES The dataset used in this study contains information on the entire adult population (ages 16 to 65)

who lived in the larger Stockholm metropolitan area between 1990 and 1998. The mean number of years of observation is 8.1. Annual demographic and socio-economic information is available for all individuals in the dataset (1.4 million individuals and 10 million person-years). Information from the National Cause of Death Register (1991-1999) is used to identify the suicide cases. I 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 1991-1996 (International Classification of Diseases, 9th revision) and X60-X84, Y87.0, or Y10-Y34 for 1997-1999 (International Classification of Diseases, 10th revision). Neighborhood-Level Data The Stockholm metropolitan area is divided into 902 so-called SAMS areas that are specifically designed to contain socially homogeneous residential areas. For the multilevel analysis, I exclude SAMS areas with fewer than 500 residents because neighborhood-level estimates based on small numbers of residents are unreliable. The multilevel analysis is based on the remaining 578 SAMS areas; the mean population size is 1,865. The neighborhood-level variables for each of the areas are calculated on the basis of all person-year data on their residents from 1990 to 1998.

NEIGHBORHOOD NETWORKS AND SUICIDE 13

Individual-Level Data I use different selection procedures to select the year of observation (year t) for those who had died from suicide. For those who committed suicide, information from the year prior to death is used.3

For those who had not committed suicide, I randomly select a one-year record of

observation (year t). This selection procedure has the advantage of reducing the computational complexity while still utilizing information on all individuals. The outcome measure is whether the person committed suicide in year t+1. The way in which the subsample is constructed means that the dependent variable in the regression analysis equals the log-odds for a person to commit suicide between 1991 and 1999. Excluding individuals with missing data on the individual-level variables (N=809, 0.06%), this study uses data on 1,384,473 individuals (691,062 men and 693,411 women). Of the total, 2,221 committed suicide (1,430 men and 791 women). Control and Explanatory Variables Table 1 presents the definitions and descriptive statistics of the neighborhood- and individuallevel variables. Age and gender are included as control variables since they are consistently associated with suicide (Hawton and Van Heeringen 2000). Neighborhood size is included to control for health-care resources. Previous studies suggest that the effects of income, unemployment and ethnicity may depend on the income and unemployment levels in the neighborhoods and their ethnic compositions (Hawton et al. 2001; Neeleman and Wessely 1999; Platt and Kreitman 1990).

3 The mean number of years of observation for those who commit suicide is 4.47, shorter than the 8.1 years of all individuals in the dataset. Thus, while the percentage of records from those committing suicides is 0.09% of the personyear level data, it is 0.16% in this person-level dataset. However, since the outcome measure is whether a person who ever lived in the Stockholm metropolitan area committed suicide during the period 1991 to 1999, the individual is the proper analytical unit and no adjustments are required.

NEIGHBORHOOD NETWORKS AND SUICIDE 14

Income is log-transformed because of its skewed distribution. An additional measure of economic deprivation, is whether an individual received social welfare benefits during year t. Whether the individual received unemployment benefits during year t is used as a proxy measure of unemployment. Single- parent status has been shown to be associated with a high suicide risk (Ringback Weitoft, Haglund, and Rosen 2000) and it is used here as a proxy measure of family decomposition. Country of birth is used as a proxy measure of ethnicity. Following Hjern and Allebeck (2002), I classify the countries of origin into six groups (See notes for Table 1). Apart from age and gender, all other individual-level variables are paired by an analogous neighborhood-level variable, e.g., logged personal income is matched by the neighborhood’s median income. ------------------------------------Table 1 Here -------------------------------------

IV.

STATISTICAL METHODS

Mapping To examine the geographic variations of suicide risk in the Stockholm metropolitan area, both standardized mortality ratios (SMRs) and Empirical Bayes standardized mortality ratios (EBSMRs, Clayton and Kaldor 1987)) of suicide are calculated for all 902 SAMS areas. The SMRs are calculated as: SMR (k) = [Observed number of suicides(k)/Expected number of suicide(k)] X 100

NEIGHBORHOOD NETWORKS AND SUICIDE 15

where the expected number of suicides are estimated by extrapolating the age- and gender-specific suicide rates of the entire greater Stockholm metropolitan population from 1991 to 1999 to the age and gender distribution in each SAMS area in 1995. However, the small population size of several SAMS areas produces extreme SMRs. Therefore, I employ the Bayesian approach to adjust for neighborhood size. In calculating the EBSMRs, the expected number of suicides is assumed to follow a Poisson distribution, and the information from the more-populated areas is “borrowed” to estimate the expected number in lesspopulated areas (Rabe-Hesketh and Skrondal 2005). The EBSMRs are estimated with the Stata program GLLAMM (Rabe-Hesketh and Skrondal 2005). After the adjustment, the extreme SMRs are “shrunk” towards the mean for the entire area. Multilevel Analysis To account for the clustered nature of the data when estimating the cross-level effects (Raudenbush and Bryk 2002), I employ multilevel logistic regression models. I use the HLM 6.02a software (Scientific Software International, Inc., Lincolnwood, IL) to fit the two-level models and to obtain restricted maximum likelihood (REML) estimates. Age, logged personal income, and all neighborhood-level variables are grand mean centered. Wald tests are used to determine levels of significance.

V.

RESULTS

Local Geographies of Suicide Within the greater Stockholm area, the range of the SMRs of suicide is 17,727. The range of the EBSMRs is 197.2 (45.3-240.5). Thus, after smoothing there is still a fivefold difference in a

NEIGHBORHOOD NETWORKS AND SUICIDE 16

neighborhood’s risk of suicide between the lowest- and highest-risk areas. This range is higher than the ranges of EBSMRs that Saunderson and Langford (1996) estimate among different population sub-groups in England (64.8-145.5). My finding suggests that despite the small size of the Stockholm metropolitan area, the sufficient variability in suicide risk across the neighborhood must be explained. ------------------------------------------Figure 4 Here -------------------------------------------

Figure 4 is a map of the EBSMRs. Breault (1994) suggests that the clear East-West divide in suicide risk in the U.S. is a major confounder that has generated spurious relationships between socioeconomic indicators and suicide rates. As Figure 4 shows, although local clusters of aboveaverage SMRs are found in adjacent SAMS areas, the Stockholm metropolitan area does not have a definitive geographical pattern as in the U.S. It also shows that although urban areas tend to have above-average EBSMRs, high suicide risk is not limited to those areas. This suggests that we cannot explain the geographical variability in suicide risk simply by the different levels of urbanity.

Multi-level analysis I now turn to the results from the multi-level analysis.4 The first three of the five multilevel models are used to gauge how much of the variance across neighborhoods can be explained by the

4 Before fitting the multilevel models, standard logistic regression models are used to explore the relationships between suicide risk and the neighborhood- and individual-level variables. The STATA program, RELOGIT (Tomz, King, and Zeng 1999) is used to examine whether the rare occurrence of suicide in the data leads to an overestimation of the

NEIGHBORHOOD NETWORKS AND SUICIDE 17

individual- and neighborhood-level predictors. The last two models test the cross-level interactions. In all models, subscripts denote the jth individual in the kth neighborhood.

Model 1. Unconditional model This is an unconditional model with only the grand mean as the predictor and a random neighborhood effect. Individual-level model: Log(Ρ jk / 1 − Ρ jk ) = β 0 k

Neighborhood-level model:

β 0 k = γ 00 + ε 0 k , where Ρ jk equals the probability that personj in the neighborhoodk committed suicide between 1991 and 1999. γ 00 is the average log-odds of suicide in the population, β 0 k is the neighborhood-specific log-odds, and ε 0 k is the unobserved error term or random effect in the neighborhood. ε 0 k is assumed to be normally distributed with zero mean and variance τ 00 . In the unconditional model, the estimated average log-odds of suicide in all areas, γ00, is 6.496 (Model 1 in Table 2). It corresponds to a probability ( ρ 00 ) = of 1/(1+exp(-6.496) =0.00151,5 or 151 suicides per 100,000 between during the 9-year study period. The estimated variance of β 0 k , τ 00 is statistically significant ( χ 2 =817.520; d.f.=577; p-value=0.000). Assuming β 0 k is

standard errors ((King and Zeng 2000; King and Zeng 2002). Using RELOGIT to estimate the same regression equations yields almost identical point estimates and standard errors, providing some support for the notion that the rare occurrence of suicide does not lead to a substantial bias in the logistic regression models. 5 This estimate is smaller than the actual proportion of the population who committed suicide between 1991 and 1999, which is 0.0016, because transforming the population-average log-odds back to a probability gives the median rather than the mean probability. The positively skewed distribution of the probabilities of suicide across neighborhoods implies that the median will be lower than the mean ((Raudenbush and Bryk 2002, p.297).

NEIGHBORHOOD NETWORKS AND SUICIDE 18

normally distributed, the 95% confidence interval for β 0 k equals [-7.171,-5.821]. In other words, this model estimates that 95% of the log-odds of suicide in the neighborhoods fall between -7.171 and 5.821, or between 77 and 296 suicides per 100,000 during the study period.

----------------------------------------------Table 2 here -----------------------------------------------

Model 2. Fixed individual-level effects The second model includes the fixed effects of all individual-level predictors. Individual-level model: Log(Ρ jk / 1 − Ρ jk ) =

β 0 k + β1k Age j + β 2 k Genderj + β3k Income j + β 4 kUnemployment benefits j + β 5 k Social welfare benefits j + β 6 k Single parent j + β 7 k Finland j + β8 kWestern j + β9 k Eastern European j + β10 k Southern European j + β11k Middle - East j + β12 k Other non - European j

Neighborhood-level model:

β 0 k = γ 00 + ε 0 k βδ k = γ δ 0

for δ >0

All of the individual-level coefficients, βδ k , for δ >0 are assumed to be invariant across the neighborhoods (i.e. all of the individual-level variables are assumed to have only a fixed effect). As shown in Table 2, including the individual-level predictors reduces the variance component only

NEIGHBORHOOD NETWORKS AND SUICIDE 19

slightly and the estimated confidence intervals for the neighborhood intercept remain more or less the same. This suggests that self-selection into different neighborhoods along the observed individual characteristics contributes relatively little to the geographical variability in suicide risk.

Model 3. Neighborhood-level effects Comparing the variance component in Model 2 with Model 3 can help to determine the relative importance of the individual-level and the neighborhood-level variables in generating the geographical variations. In Model 3, the individual-level model has only the grand intercept as the predictor (as in Model 1). The neighborhood-level model is formulated as: Neighborhood-level model: β0k =

γ00 + γ01 Population sizek + γ02 Median incomek + γ03 % Unemployment benefitsk + γ04 % Social welfare benefitsk + γ05 % Single parentsk + γ06 % Finlandk + γ07 %Westernk + γ08 % Eastern Europeank + γ09 % Southern Europeank + γ010% Middle Eastk + γ011 % Other non-Europeank + ε0k The neighborhood intercept β0k is allowed to vary with the neighborhood-level variables (γ01-

γ011) in addition to the random component (ε0k). Controlling for the neighborhood-level predictors, the variance component ( τ 00 ) estimated by Model 3 is smaller than Model 2. Now the estimated suicide rates in different neighborhoods only range from 105 per 100,000 to 176 per 100,000. In other words, including the neighborhood-level variables reduces more of the estimated variability in the neighborhood-specific suicide risk than the individual-level variables. These results suggest that the neighborhood-level predictors are more important in explaining the variance in suicide risk across the neighborhoods.

Model 4. Full model

NEIGHBORHOOD NETWORKS AND SUICIDE 20

Model 4 examines the effects of neighborhood- and individual-level predictors simultaneously and tests whether the hypothesized negative cross-level interactions are statistically significant. It includes all of the individual- and neighborhood-level predictors used in Model 2 and Model 3. The level-1 slopes β3k - β12k are now allowed to vary by the analogous neighborhood-level variables and have a random component. Including the individual-, neighborhood- and cross-level effects further reduces the variance component τ 00 . Now the estimated suicide rates across the neighborhoods only range from 114 to 118 per 100,000. The χ 2 statistic for τ 00 also becomes insignificant (p-value>0.5), suggesting that the parameters in Model 4 explain a large part of the variability in suicide risk across the neighborhoods. Table 3 reports the parameter estimates of Model 4. Controlling for the effects of the individual-level variables, the following neighborhood-level variables are significant at the 0.05 level: (1) median income; (2) the percentage of residents who receive unemployment benefits; (3) the percentage of residents who received social welfare benefits; (4) the percentage of residents born in Southern Europe; and (5) the percentage of residents born in other non-European countries. Notably, seven out of the ten neighborhood variables have opposite effects at the individual-level. For example, being born in Southern Europe is associated with a significantly lower risk of suicide compared to the Swedes, while the percentage of residents in the neighborhood born in Southern Europe is associated with a significantly higher risk. Differences in the directions of association between levels are also found for median income, unemployment benefits, being a single parent, and being born in a Western country, Eastern Europe or the Middle East. Apart from the cross-level interaction concerning Eastern European birth, all of the remaining cross-level interactions tested have a negative sign as predicted. Thus the results generally support the

NEIGHBORHOOD NETWORKS AND SUICIDE 21

mechanisms presented in Section II. The following cross-level interactions are significant at the 0.05 level in both the one- and the two-tailed tests: (1) personal income and median income in the neighborhood; (2) received social welfare benefits and the percentage of neighbors who received them; (3) born in Southern Europe and the percentage of residents born in Southern Europe; and (4) born in another non-European country and the percentage of residents born in other nonEuropean countries. ----------------------------------------------Table 3 Here -----------------------------------------------

Model 5. Reduced model To reduce the number of redundant parameters, Model 5 retains all significant variables in Model 4 and the neighborhood- and individual-level variables having significant cross-level interactions. I estimate only the random component of the neighborhood intercept (ε0k) and the random components of the slopes of the individual variables that show significant cross-level effects. Table 4 reports the estimates of Model 5. The results are largely similar to those reported in Table 3, except that the neighborhood-level effect of the percentage of residents born in other nonEuropean countries has become insignificant.

----------------------------------------------Table 4 Here -----------------------------------------------

NEIGHBORHOOD NETWORKS AND SUICIDE 22

Although the cross-level interactions concerning social welfare benefits and born in other non-European countries are still statistically significant, it can be argued that their effect sizes are rather small: the odds ratio for receiving social welfare benefits by the percentage who received them is 0.97, while it is 0.94 for born in another non-European country by the percentage born in other non-European countries. In contrast, the cross-level interaction effects of income (O.R. = 0.57) and born in Southern Europe (O.R. = 0.79) are substantial. Figure 5 illustrates the cross-level effects concerning income and being born in Southern Europe. As shown in Figure 5a, the suicide risk for those with a low personal income (25th percentile) appears to be slightly lower than those with a high personal income (75th percentile) when the median income in the neighborhood is low. However, the suicide risk relative to the high income group increases when the income level of the neighborhood increases; it becomes higher than the high income group when the neighborhood income reaches the level of the overall median income. Such a pattern is similar to the pattern depicted in Figure 1c. In comparison, the effect of being born in Southern Europe (Figure 5b) matches Figure 1b. In neighborhoods that contain few people born in Southern Europe, the suicide risk for those born there is still lower than for those born in other countries. However, the relative suicide risk decreases substantially when the percentage of neighbors from Southern Europe increases. Moreover, the suicide risk of those born in other countries increases with the percentage of residents born in Southern Europe, which is also consistent with the proposed mechanisms.

----------------------------------------------Figure 5 Here -----------------------------------------------

NEIGHBORHOOD NETWORKS AND SUICIDE 23

VI.

CONCLUSIONS This paper examines how contextual interactions in neighborhoods can contribute to

geographical patterns of suicide. I argue that the contextual interaction mechanisms implicit in the classic theories of suicide (social support, normative pressure and dissonance-related stress) predict that the effect of an individual attribute on his/her suicide risk should depend on the neighborhood context. That is, the more common the attribute, the lower the associated suicide risk. The results from multilevel analysis support the prediction. Neighborhood interactions do matter, and having neighbors similar to oneself is generally protective against suicide. As Section II illustrates, the proposed mechanisms imply negative interactions between an individual-level variable and its analogous measure at the neighborhood-level in multilevel analysis. As predicted, all ten cross-level interactions with one exception have a negative sign. Statistically significant cross-level interactions are found for income, receiving social welfare benefits and two of the ethnic groups. The cross-level effects of income and born in Southern Europe are substantial. These results suggest that at least some of the geographical variability in suicide rates is the result of the interactions between neighborhood compositions and individual characteristics. Among the three measures of socio-economic deprivation, income shows the most substantial cross-level effect, while receiving social welfare benefits has a significant, though smaller, effect. The insignificant result of receiving unemployment benefits is most likely because it is a poor measure of unemployment (not all the unemployed would have received the benefits). This also explains why receiving unemployment benefits is not significantly associated with suicide risk even at the individual-level in this dataset. Although the cross-level interaction between being a single parent and its prevalence in the neighborhood has the predicted negative sign, it is not statistically significant. Therefore, there is no conclusive evidence for a cross-level effect in relation to family integration. In contrast, two of the

NEIGHBORHOOD NETWORKS AND SUICIDE 24

six ethnicities display significant cross-level interactions. Although some caution is necessary given the number of ethnic groups included, it provides further evidence of the cross-level interaction effects reported by Neeleman and Wessely (1999). Two previous studies find no strong support for cross-level effects when considering socioeconomic variables (Agerbo, Sterne, and Gunnell 2007; Martikainen, Mäki, and Blomgren 2004b). Given the differences in data and statistical methods (neither study employs random effects models to estimate the cross-level interactions), it is difficult to explain the varied results. However, some of the results from these two studies support that the suicide risk associated with certain socioeconomic variables differs, albeit modestly, with neighborhood characteristics. van Tubergen et al. (2005) also find no support for the cross-level interactions. Nonetheless, all of these previous studies use much larger geographical units than the SAMS areas. The mean population size is approximately 20,000 in Agerbo et al. (2007), 59,000 in Martikainen et al. (2004b) and 11,500 in van Tubergen et al. (2005). As mentioned before, mechanisms of neighborhood interactions are much less likely to span across large, heterogeneous areas. The use of a more appropriate geographical unit may explain why this study finds more support for contextual interactions. Social support, normative pressure and dissonance-related stress are proposed as the underlying mechanisms the current findings. Certainly, competing hypotheses can be formulated: Durkheim claimed that high levels of social integration/regulation also are associated with high suicide rates. Thus, at the other end of the integration/regulation continuum, the cross-level effects can be positive. Similarly, the “double deprivation” hypothesis (Wagstaff and van Doorslaer 2000) states that the socio-economically deprived will suffer most when they also live in deprived neighborhoods. For the Stockholm dataset, I do not find empirical support for such positive crosslevel interaction effects. Yet it remains possible for such positive effects in societies with higher levels of social integration/regulation.

NEIGHBORHOOD NETWORKS AND SUICIDE 25

Four neighborhood-level variables (i.e., median income, percentages who received unemployment benefits and social welfare benefits, and percentage born in Southern Europe) remain statistically significant when the analogous individual-level variables are controlled for. Several possibilities may account for such ecological associations. First, the neighborhood-level variables may represent shared environmental characteristics that increase suicide risk, such as inadequate health-care facilities. It is also possible that these associations are due to confounding between the neighborhood variables and unobserved individual characteristics (see Oakes 2004), which causes spurious relationships between neighborhood characteristics and suicide. The potential selection effects due to omitted variables hamper the ability to draw any definitive conclusions about the neighborhood-level relationships.6 Interestingly, many effects of the explanatory variables in this paper run in opposite directions to the individual- and neighborhood-levels. Such findings clearly demonstrate the danger of ecological (Robinson 1950) and atomic fallacies (Diez Roux 2002): the problem of erroneously inferring relationship at the group level to the individual level and vice versa. The current results cast serious doubt to the practice of using neighborhood-level measures as proxies of individual-level measures when the latter are unavailable. This problem of inference across levels certainly contributes to the inconsistent results for the social correlates of suicide in addition to the effects of the contextual interactions. Disentangling the possible causes for geographical patterns of suicide rates has practical implications for policies designed to reduce health disparities. As Gunnell and Frankel (1994) point out, progress in suicide prevention is greatly constrained by the lack of evaluations, and populationwide or community-based programs are the most difficult to evaluate (see also Mann, Apter,

6

One can possibly attempt to limit the effect of omitted variables by including more individual variables, but this may produce other problems such as including intermediate variables as control variables ((Diez Roux 2001; Weinberg 1993). Hence, this paper limits the variables to those related to the cross-level interactions of interest.

NEIGHBORHOOD NETWORKS AND SUICIDE 26

Bertolote, Beautrais, Currier, Haas, Hegerl, Lonnqvist, Malone, Marusic, Mehlum, Patton, Phillips, Rutz, Rihmer, Schmidtke, Shaffer, Silverman, Takahashi, Varnik, Wasserman, Yip, and Hendin 2005). This makes research on the relevant social interaction processes especially valuable. For example, some national strategies propose to strengthen the ethnic minority communities (e.g., New Zealand's Associate Minister of Health 2006). The current findings suggest that such programs should have a positive impact, but that their effects upon other social groups should also be evaluated. This study has the following limitations. First, although the SAMS areas are designed to represent homogeneous residential areas, they only loosely resemble the actual fields of social interaction patterns. Thus the social interaction effects reported may be underestimated. Second, there is no information available about the suicides committed by the individuals who move outside the Stockholm metropolitan area during the study period. Third, the interaction processes cannot be measured directly. Further research that can directly assess the effect of social contexts on individual beliefs and desires about suicide is needed.

NEIGHBORHOOD NETWORKS AND SUICIDE 27

References

Agerbo, E., J.A. Sterne, and D.J. Gunnell. 2007. "Combining individual and ecological data to determine compositional and contextual socio-economic risk factors for suicide." Social Science and Medicine 64:451-61. Associate Minister of Health. 2006. The New Zealand Suicide Prevention Strategy 2006–2016. Wellington: Ministry of Health. Bearman, P.S. and J. Moody. 2004. "Suicide and friendships among American adolescents." Am J Public Health 94:89-95. Breault, K.D. 1994. "Was Durkheim right? A critical survey of the empirical literature on Le Suicide." Pp. 1129 in Emile Durkheim : le suicide, one hundred years later, edited by D. Lester. Philadelphia: Charles Press. Clayton, D. and J. Kaldor. 1987. "Empirical Bayes estimates of age-standardized relative risks for use in disease mapping." Biometrics 43:671-81. Diez Roux, A.V. 2001. "Investigating neighborhood and area effects on health." American Journal of Public Health 91:1783-9. —. 2002. "A glossary for multilevel analysis." Journal of Epidemiology and Community Health 56:588-94. Durkheim, E. 1897[1951]. Suicide : A Study in Sociology. Translated by G. Simpson. London: Routledge & Kegan Paul. Ellison, C.G., J.A. Burr, and P.L. McCall. 1997. "Religious Homogeneity and Metropolitan Suicide Rates." Social Forces 76:273-299. Gibbs, J.P. and W.T. Martin. 1964. Status integration and suicide : a sociological study. Eugene, Or: University of Oregon Books. Gunnell, D. and S. Frankel. 1994. "Prevention of suicide: aspirations and evidence." Bmj 308:1227-33. Hawton, K., L. Harriss, K. Hodder, S. Simkin, and D. Gunnell. 2001. "The influence of the economic and social environment on deliberate self-harm and suicide: an ecological and person-based study." Psychological Medicine 31:827-36. Hawton, K. and K. Van Heeringen. 2000. The International Handbook of Suicide and Attempted Suicide. Chichester: John Wiley & Sons. Hedström, P., Liu Ka-yuet, and M. Nordvik. Forthcoming. "Interaction domains and suicides: A populationbased panel study of suicides in Stockholm, 1991-1999." Social Forces. Heikkinen, M., H. Aro, and J. Lönnqvist. 1994. "Recent life events, social support and suicide." Acta psychiatrica Scandinavica. Supplementum 377:65-72. Hjern, A. and P. Allebeck. 2002. "Suicide in first- and second-generation immigrants in Sweden: a comparative study." Social Psychiatry and Psychiatric Epidemiology 37:423-9. King, G. and L. Zeng. 2000. "Logisitc Regression in Rare Events Data." Political Analysis 9:137-163. —. 2002. "Estimating risk and rate levels, ratios and differences in case-control studies." Statistics in Medicine 21:1409-27. Linehan, M.M., J.L. Goodstein, S.L. Nielsen, and J.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 51:276-86. Liu, K.Y., E.Y. Chen, C.L. Chan, D.T. Lee, Y.W. Law, Y. Conwell, and P.S. Yip. 2006. "Socio-economic and psychological correlates of suicidality among Hong Kong working-age adults: results from a population-based survey." Psychological Medicine 36:1759-67. Mäkinen, I. 1997. "Are there social correlates to suicide?" Soc Sci Med 44:1919-1929. Malone, K.M., M.A. Oquendo, G.L. Haas, S.P. Ellis, S. Li, and J.J. Mann. 2000. "Protective factors against suicidal acts in major depression: reasons for living." Am J Psychiatry 157:1084-8.

NEIGHBORHOOD NETWORKS AND SUICIDE 28

Mann, J.J., A. Apter, J. Bertolote, A. Beautrais, D. Currier, A. Haas, U. Hegerl, J. Lonnqvist, K. Malone, A. Marusic, L. Mehlum, G. Patton, M. Phillips, W. Rutz, Z. Rihmer, A. Schmidtke, D. Shaffer, M. Silverman, Y. Takahashi, A. Varnik, D. Wasserman, P. Yip, and H. Hendin. 2005. "Suicide prevention strategies: a systematic review." Jama 294:2064-74. Manski, C.F. 1993. "Identification of Endogenous Social Effects: The Reflection Problem." Review of Economic Studies 60:531-42. —. 2000. "Economic Analysis of Social Interactions." Journal of Economic Perspectives 14:115-36. Martikainen, P., N. Mäki, and J. Blomgren. 2004a. "The Effects of Area and Individual Social Characteristics on Suicide Risk: A Multilevel Study of Relative Contribution and Effect Modification." European Journal of Population 20:323–350. —. 2004b. "The Effects of Area and Individual Social Characteristics on Suicide Risk: A Multilevel Study of Relative Contribution and Effect Modification." European Journal of Population 20:323–350. McPherson, M., L. Smith-Lovin, and J.M. Cook. 2001. "Birds of a Feather: Homophily in Social Networks." Annual Review of Sociology 27:415-444. Moksony, F. 1994. "The whole, its parts and the level of analysis: Durkheim and the macrosociological study of suicide." Pp. 101-114 in Emile Durkheim : Le Suicide, One Hundred Years Later, edited by D. Lester. Philadelphia: Charles Press. Neeleman, J. and S. Wessely. 1999. "Ethnic minority suicide: a small area geographical study in south London." Psychological Medicine 29:429-36. Oakes, J.M. 2004. "The (mis)estimation of neighborhood effects: causal inference for a practicable social epidemiology." Social Science and Medicine 58:1929-52. Pescosolido, B.A. 1990. "The Social Context of Religious Integration and Suicide: Pursuing the Network Explanation." The Sociological Quarterly, 1990, 31, 3, fall, 337-357. —. 1994. "Bringing Durkheim into the Twenty-First century: a network approach to unresolved issues in the sociology of suicide." Pp. 264-96 in Emile Durkheim : Le suicide, One Hundred Years Later, edited by D. Lester. Philadelphia: Charles Press. Pescosolido, B.A. and S. Georgianna. 1989. "Durkheim, Suicide, and Religion: Toward a Network Theory of Suicide." American Sociological Review 54:33-48. Platt, S. and N. Kreitman. 1990. "Long term trends in parasuicide and unemployment in Edinburgh, 196887." Social Psychiatry and Psychiatric Epidemiology 25:56-61. Rabe-Hesketh, S. and A. Skrondal. 2005. Multilevel and Longitudinal Modeling using Stata. College Station, TX: : Stata Press. Raudenbush, S.W. and A.S. Bryk. 2002. Hierarchical linear models : applications and data analysis methods. Thousand Oaks: Sage Publications. Rehkopf, D.H. and S.L. Buka. 2006. "The association between suicide and the socio-economic characteristics of geographical areas: a systematic review." Psychological Medicine 36:145-57. Ringback Weitoft, G., B. Haglund, and M. Rosen. 2000. "Mortality among lone mothers in Sweden: a population study." Lancet 355:1215-9. Robinson, W.S. 1950. "Ecological correlations and the behavior of individuals." American Sociological Review 15:351-7. Rodgers, G.B. 2002. "Income and inequality as detriments of mortality: An. international cross-section analysis." Population Studies 33: 343-351. Sainsbury, P., J. Jenkins, and A. Levey. 1980. "The social correlates of suicide in Europe." Pp. 38-53 in The Suicide Syndrome, edited by E. Farmer and S. Hirsch. London: Croom Helm. Sampson, R.J., J.D. Morenoff, and T. Gannon-Rowley. 2002. "Assessing 'neighborhood effects': Social processes and new directions in research." Annual Review of Sociology 28:443. Saunderson, T.R. and I.H. Langford. 1996. "A study of the geographical distribution of suicide rates in England and Wales 1989-92 using empirical bayes estimates." Soc Sci Med 43:489-502. Thorlindsson, T. and T. Bjarnason. 1998. "Modeling Durkheim on the micro level: A study of youth suicidality." American Sociological Review 63:94-110. Tomz, M., G. King, and L. Zeng. 1999. "RELOGIT: Rare Events Logistic Regression Version 1.1." Cambridge, MA: Harvard University.

NEIGHBORHOOD NETWORKS AND SUICIDE 29

van Tubergen, F., M. te Grotenhuis, and W. Ultee. 2005. "Denomination, Religious Context, and Suicide: Neo-Durkheimian Multilevel Explanations Tested with Individual and Contextual Data." American Journal of Sociology 111:797. Wagstaff, A. and E. van Doorslaer. 2000. "Income inequality and health: what does the literature tell us?" Annual Review of Public Health 21:543-67. Weinberg, C.R. 1993. "Toward a clearer definition of confounding." American Journal of Epidemiology 137:1-8.

A person without X % of neighbours with X (a) X = Risk factor

A person without X

A person with X

% of neighbours with X (b) X = Protective factor

Individual suicide risk

A person with X

Individual suicide risk

Individual suicide risk

NEIGHBORHOOD NETWORKS AND SUICIDE 30

A person with X

A person without X

% of neighbours with X (c) X = Risk OR protective factor

Figure 1. Negative cross-level interactions depicted at the individual-level

% of residents with X (a) X = Risk factor at the individual-level

Neighborhood suicide rate

Neighborhood suicide rate

Neighborhood suicide rate

NEIGHBORHOOD NETWORKS AND SUICIDE 31

% of residents with X (b) X = Protective factor at the individual-level

% of residents with X (c) X = Risk OR protective factor at the individual-level

Figure 2. Negative cross-level interactions depicted at the neighborhood-level.

A person with X

A person without X

% of neighbours with X (a) No diminishing benefits

Individual suicide risk

Individual suicide risk

NEIGHBORHOOD NETWORKS AND SUICIDE 32

A person with X

A person without X

% of neighbours with X (b) Diminishing benefits

Figure 3. Saturation of benefits of having yet another similar neighbor. In (a) the benefits of having a similar neighbor in terms of the reduction in suicide risk do not change with the proportion of similar neighbors in the neighborhood. In (b) such benefits decrease with the proportion of such neighbors in the neighborhood.

NEIGHBORHOOD NETWORKS AND SUICIDE 33

Figure 4. Empirical Bayes Estimated Standard Mortality Ratio in Stockholm Metropolitan Area, 1991-1999

NEIGHBORHOOD NETWORKS AND SUICIDE 34

Table 1. Individual-level and neighborhood-level variables Variables

Individual-level Definition

Age Gender Neighborhood size Income

Age at year t Male=0, Female=1 -

Unemployment benefits Social welfare benefits Single parent

Received unemployment benefits (0=No; 1=Yes) Received social welfare benefits (0=No; 1=Yes) Being a single parent (0=No; 1=Yes)

Country of birthc Finland

Logged disposable incomeb

Neighborhood-levela Mean Definition (S.D.) Number of adults ages 16-65 1864.69 years in the neighborhood (1782.04) Logged median disposable 1.12 income2 of the neighborhood (0.12) % of residents received 8.32 unemployment benefits (3.14) % of residents received social 7.00 welfare benefits (6.16) % of residents who were single 6.34 parent (2.66)

Born in Finland (0=No; 1=Yes) Born in Western country (0=No; 1=Yes)

% of residents born in Finland

Eastern Europeane

Range 503.003652.92 0.58-1.43 2.68-17.70 0.55-39.53 1.76-16.76 0.97-22.12

% of residents born in Western countries

5.68 (3.21) 2.89 (1.39)

Born in Eastern European country (0=No; 1=Yes)

% of residents born in Eastern European countries

1.82 (1.03)

0.03-6.45

Southern Europeanf

Born in Southern European country (0=No; 1=Yes)

% of residents born in Southern European countries

1.48 (1.59)

0.00-13.94

Middle Eastg

Born in Middle East country (0=No; 1=Yes) Born in other non-European country (0=No; 1=Yes)

% of residents born in Middle East countries % of residents born in other non-European countries

2.89 (5.03) 3.17 (3.35)

0.00-40.26

Westernd

Other nonEuropeanh a. b. c. d. e. f. g. h.

0.94-26.13

0.08-23.72

Calculated based on all 1990-1998 person-year records of those ages 16 to 65 in the 578 SAMS areas with 500 or more inhabitants. In base amount. One base-amount equals approximately 33,000 Swedish Krona Reference category: born in Sweden Denmark, Norway, Iceland, Germany, Belgium, France, Luxemburg, the Netherlands, Austria, Great Britain, Ireland, Andorra, Lichtenstein, the Vatican, Switzerland, Monaco, Malta, San Marino, United States, Canada, Australia, New Zealand and other Oceania countries Estonia, Lithuania, Albania, Bulgaria, Romania, Czechoslovakia, Czech Republic, Slovakia, Hungary, Poland, Moldavia, Russia, Republic of Belarus and Ukraine Italy, Spain, Portugal Greece Bosnia-Herzegovina, Croatia, Macedonia and Slovenia Egypt, Libya, Tunisia, Algeria, Morocco, Israel, Palestine, Syria, Lebanon, Jordan, Yemen (south), Yemen, United Arab Emirates, Kuwait, Bahrain, Qatar, Saudi Arabia, Iran, Iraq and Turkey Asian (including Japan), African, Central American and South American countries

NEIGHBORHOOD NETWORKS AND SUICIDE 35

Table 2. Estimated logged odds of suicide and their confidence intervals from multilevel models Model 1: Unconditional

Model 2: Fixed individuallevel variables

Model 3: Neighbor-hoodlevel variables

Model 4: Full model

-6.496

-6.431

-6.339

-6.524

0.00151

0.00161

0.00136

0.00032

0.119 817.520 (577) 0.000

0.094 755.857 (577) 0.000

0.017 631.296 (565) 0.027

0.016 523.797 (532) >0.50

Lower 95% C.I. of β 0k

-7.171

-7.030

-6.856

-6.772

Upper 95% C.I. of β 0k

-5.821

-5.831

-6.340

-6.275

Lower 95% C.I. of ρ 0k

0.00077

0.00088

0.00105

0.00114

Upper 95% C.I. of ρ 0k

0.00296

0.00293

0.00176

0.00188

Grand intercept

γ 00 ρ 00

Variance component

τ 00 χ 2 for τ00, (d.f.) p-value for τ00 Range of neighborhood intercepts

NEIGHBORHOOD NETWORKS AND SUICIDE 36

Table 3. Model 4: Multilevel model with all individual-, neighborhood- and cross-level predictors β (S.E.) Neighborhood-level effects Grand intercept Population size Median income % Unemployment benefits % Social welfare benefits % Single parent Ethnicity % Finland % Western % Eastern European % Southern European % Middle East %Other non-European Individual-level effects Age Female Income Unemployment benefits Social welfare benefits Single parent Ethnicity Finland Western Eastern European Southern European Middle East Other non-European

OR (95%C.I.)

p-value

-6.524 5.E-06 1.903 0.088 0.021 0.025

(0.038) (9.E-06) (0.411) (0.014) (0.011) (0.016)

0.002 (0.001,0.002) 1.000 (1.000,1.000) 6.703 (2.991,15.021) 1.092 (1.064,1.122) 1.021 (1.000,1.043) 1.026 (0.994,1.058)

0.000*** 0.555 0.000*** 0.000*** 0.048* 0.112

0.000 0.006 0.021 0.056 0.009 -0.025

(0.010) (0.031) (0.033) (0.016) (0.008) (0.013)

1.000 (0.981,1.020) 1.006 (0.948,1.069) 1.021 (0.957,1.090) 1.058 (1.024,1.093) 1.010 (0.995,1.025) 0.975 (0.950,1.001)

0.965 0.833 0.525 0.001** 0.212 0.058

0.029 -0.605 -0.038 -0.037 1.447 -0.152

(0.002) (0.046) (0.039) (0.095) (0.068) (0.113)

1.029 (1.026,1.032) 0.546 (0.499,0.597) 0.962 (0.892,1.038) 0.963 (0.799,1.162) 4.251 (3.718,4.861) 0.859 (0.688,1.071)

0.000*** 0.000*** 0.322 0.695 0.000*** 0.178

0.284 -0.250 -0.230 -0.421 -1.208 -0.829

(0.084) (0.134) (0.180) (0.201) (0.199) (0.180)

1.329 (1.127,1.568) 0.779 (0.599,1.013) 0.795 (0.559,1.130) 0.656 (0.443,0.973) 0.299 (0.202,0.441) 0.436 (0.307,0.621)

0.001** 0.062 0.201 0.036* 0.000* 0.000*

0.967 (0.901,1.038) 0.546 (0.326,0.912) 0.986 (0.936,1.038)

0.353 0.021* 0.582

0.973 (0.960,0.987)

0.000***

0.990 (0.953,1.028) 0.975 (0.778,1.221) 1.121 (0.899,1.398) 0.787 (0.668,0.926) 0.993 (0.967,1.020) 0.933 (0.883,0.986)

0.592 0.825 0.311 0.005* 0.626 0.014*

Cross-level effects Single parent by % being single parent -0.033 (0.036) Personal income by median income -0.606 (0.262) Received unemployment benefits by % -0.014 (0.026) received unemployment benefits Received social welfare benefits by % -0.027 (0.007) received social welfare benefits Ethnicity Finland*% Finland -0.010 (0.019) Western*% Western -0.025 (0.115) Eastern Europe*% Eastern Europe 0.114 (0.113) Southern Europe*% Southern Europe -0.240 (0.083) Middle East*% Middle East -0.007 (0.014) Other non-European*% Other non-0.069 (0.028) European Number of neighborhoods=578; Number of individuals = 1,384,473 Two-tailed test: * p<.05, **p<.01, ***p<.001

NEIGHBORHOOD NETWORKS AND SUICIDE 37

Table 2. Model 5: Multilevel model with selected individual-, neighborhood- and crosseffects predictors Β

(S.E.)

OR (95%C.I.)

Neighborhood-level effects Grand intercept Median income % Unemployment benefits % Social welfare benefits % Southern European %Other non-European

-6.519 1.863 0.090 0.030 0.060 -0.018

(0.035) (0.331) (0.012) (0.009) (0.016) (0.012)

0.001 (0.001,0.002) 6.441 (3.365,12.331) 1.094 (1.068,1.120) 1.031 (1.013,1.049) 1.062 (1.029,1.096) 0.982 (0.960,1.005)

0.000*** 0.000*** 0.000*** 0.001** 0.000*** 0.132

Individual-level effects Age Female Income Unemployment benefits Social welfare benefits Finland Southern European Middle East Other non-European

0.028 -0.623 -0.029 -0.071 1.452 0.278 -0.383 -1.237 -0.811

(0.002) (0.044) (0.038) (0.074) (0.066) (0.074) (0.194) (0.149) (0.176)

1.029 (1.026,1.032) 0.537 (0.492,0.585) 0.971 (0.902,1.046) 0.932 (0.807,1.077) 4.271 (3.755,4.858) 1.321 (1.142,1.529) 0.682 (0.466,0.997) 0.290 (0.217,0.389) 0.444 (0.314,0.628)

0.000*** 0.000*** 0.447 0.339 0.000*** 0.000*** 0.048* 0.000*** 0.000***

-0.560 (0.255) -0.029 (0.007)

0.571 (0.347,0.941) 0.971 (0.958,0.985)

0.028*

-0.236 (0.080) -0.065 (0.027)

0.790 (0.674,0.924) 0.937 (0.888,0.987)

Cross-level effects Personal income*Median income Received social welfare benefits by % received social welfare benefits Southern Europe* % Southern Europe Other non-European* % Other nonEuropean

Number of neighborhoods=578; Number of individuals = 1,384,473 Two-tailed test: * p<.05, **p<.01, ***p<.001

p-value

0.000*** 0.004* 0.016*

NEIGHBORHOOD NETWORKS AND SUICIDE 38

0.0020

25th percentile of logged personal income 75th percentile of logged personal income Probability of Suicide

Probability of suicide

0.0020

0.0015

0.0010

0.0005

0 -0.15

-0.07

0.00

0.07

0.15

(a) Logged neighborhood’s median income, grand mean centered

Born outside of Southern Europe Born in Southern Europe

0.0015

0.0010

0.0005

0 -1.23

-0.56

0.12

0.79

1.47

(b) % born in Southern Europe, grand mean centered

Figure 5. Cross-level effects of income and born in Southern Europe