Abstract
We aimed to demonstrate the utility of an item-level network analysis approach to suicide risk by testing the interpersonal psychological theory of suicide (IPTS) among 402 psychiatric inpatients. We hypothesized that specific thwarted belongingness (TB) or perceived burdensomeness (PB; Interpersonal Needs Questionnaire items) facets would positively relate to passive or active suicide ideation and that these facets would positively relate to each other and form distinct clusters. We also tested TB and PB facets central to the networks as predictors of suicide ideation compared with the full TB and PB subscales. Face-valid items congruent with latent constructs proposed by the IPTS (i.e., feelings of burden on society, feeling that one does not belong) were the only two facets uniquely predictive of passive and active suicide ideation. Facets of TB and PB did not form distinct clusters. Item-level network analysis may have important conceptual, assessment, predictive, and clinical implications for understanding suicide risk.
Keywords
Suicide is the 10th leading cause of death in the United States; there were approximately 15 suicides per 100,0000 individuals as of 2018 (National Institute of Mental Health, 2021). As the suicide rate continues to rise (National Institute of Mental Health, 2021), we as researchers need to turn our attention to high-risk populations that may be driving this increase. Psychiatric inpatients are at high risk for suicide; there are an estimated 101 deaths by suicide per 100,000 psychiatric hospital admissions (Neuner et al., 2008). A recent meta-analysis suggests that many known suicide risk factors, which have been frequently tested in isolation or within relatively simple models, fail to predict suicide ideation and suicidal behaviors better than chance (Franklin et al., 2017). Thus, current theoretical conceptualizations of suicide risk and/or the statistical methods used to test suicide risk may be impeding the ability to predict, prevent, and manage suicide risk. Novel frameworks and statistical approaches, such as network analysis, may better clarify suicide risk among high-risk psychiatric inpatients.
Network analysis is a statistical approach that conceptualizes latent constructs as a complex network of relations among interconnected facets. Networks are graphical representations of facets (i.e., nodes; represented as circles) and relations among two or more facets (i.e., edges; represented as lines connecting circles; a more specific description of network analysis is provided in the Method section) in which the direct and indirect relations among facets are interpreted as genuine rather than spurious. Partial networks use a matrix of conditional relations, which provide a quantifiable, graphical display of each facet’s unique contributions within the network. When applied to clinical psychology, these networks have important implications for how psychological phenomena is conceptualized, such that the constellation of symptoms or experiences (facets) and their unique associations (edges) constitute the psychological disorder or experience itself and thus should be the focus for understanding potential mechanistic processes that may serve as important treatment targets and directly affect clinical outcomes.
Network analysis has been used for conceptualizing various psychological phenomena (Borsboom & Cramer, 2013; Cramer et al., 2012; Frewen et al., 2013; McNally et al., 2015; Robinaugh et al., 2014; Ruzzano et al., 2015). More recently, it has been used for conceptualizing suicide risk (de Beurs et al., 2017, 2019). From a mereological network analysis approach, risk for suicide ideation is conceptualized as a system of interconnected or interrelated clusters of facets (i.e., risk factors) in which certain facets and relations emerge as more influential or central relative to others. Network analysis may allow researchers to better examine and refine complex models of suicide risk by estimating the relations between various risk factors as well as their relations with suicide-related outcomes (e.g., suicide ideation). Understanding risk factors in the context of a myriad of other risk factors increases the ecological and clinical utility of models of suicide risk.
Network analysis has only begun to be used to test theoretical models of suicide risk, including the interpersonal-psychological theory of suicide (IPTS; de Beurs et al., 2019; Joiner, 2005; Van Orden et al., 2010). According to the IPTS, the presence of either thwarted belonging (TB; indicated by perceived disconnection from others and a lack of reciprocal caring relationships) or perceived burden (PB; indicated by self-hatred and the belief that one’s death is worth more to others than one’s life) increases the risk for passive suicide ideation (i.e., desire for death). However, when TB and PB (i.e., thwarted interpersonal needs) are experienced simultaneously and the individual feels hopeless about these states changing, active suicide ideation is posited to occur. Meta-analytic findings from a decade of research generally support the IPTS; however, most of the studies have been conducted in nonclinical samples and have used latent constructs (Chu et al., 2017). This has prevented researchers from understanding how specific facets of the IPTS do or do not apply to high-risk clinical populations, such as psychiatric inpatients, slowing the progress of practical application.
De Beurs et al. (2019) provided preliminary evidence for the utility of a theory-based network analysis of suicide risk using a nonclinical sample of young adults (18–34 years old) in Scotland. They found that in a model that included only TB and PB and suicide ideation, both TB and PB were directly associated with suicide ideation. However, when modeled in larger networks that included additional theoretical risk factors, PB, but not TB, was directly associated with suicide ideation. The differences between these networks echo the inconsistent findings in the literature regarding TB and PB’s relative importance as they relate to suicide ideation (Chu et al., 2017; Ma et al., 2016). These inconsistent findings may reflect issues with the current conceptualization of the IPTS and the reliance on latent-variable models used to test the IPTS (e.g., masked effects in multivariate models as a result of highly correlated predictors; Mitchell et al., 2017). The use of latent factors of TB and PB within the overall network by de Beurs et al. (2019) and the use of suicide ideation, rather than distinguishing between passive and active suicide ideation, may limit their network’s utility to test the specific postulations of the IPTS. In the current study, we aimed to address these limitations by estimating an item-level network analysis of suicide risk. This provides for a fine-grained examination of the unique associations between facets of TB, PB, and passive and active suicide ideation, which may directly affect evidence-based practice in suicide risk assessment and management by providing more specific assessment and intervention targets for clinicians. Furthermore, we posited that an item-level, theory-driven network analysis approach of suicide risk may allow for a more accurate examination of complex suicide-risk models, which may clarify existing literature and refine theories of suicide risk.
An item-level network analysis of theoretical risk factors may also allow for a more complete understanding of the complex relations between risk factors and suicide ideation than latent variable methods. Issues related to latent-variable models may be particularly relevant to testing the IPTS, which has been primarily tested within a latent variable approach. Conceptually, latent-variable models assume local independence, meaning the indicators (i.e., assessment items) are assumed to be unrelated beyond their association with the latent construct (e.g., Borsboom, 2008). This may be problematic for indicators of PB (e.g., thoughts about being a burden on society, thoughts such as “people would be happier without me”) and TB (e.g., people feeling as if they do not belong, rarely interacting with people who care about them) as assessed by the 15-item Interpersonal Needs Questionnaire (INQ; Van Orden et al., 2012). It is plausible that the respective indicators are reciprocal. For example, if people feel like they do not belong, that may lead them to have fewer interactions with people who care about them, which, in turn, may further increase their feelings of not belonging. Likewise, people who feel that others would be better off if they were gone or that others would be happier without them may generalize to more global beliefs that they are a burden on society. From a network analysis perspective, the reciprocal relations between individual facets of TB and PB constitute the broader constructs of TB and PB and do not arise because of some unobserved latent construct or cause the latent construct. Thus, the direct and indirect relations among facets of TB, PB, and suicide-related outcomes are modeled and interpreted as genuine rather than spurious, and the relations are viewed as clinically meaningful.
The nuanced perspective obtained from an item-level network analysis approach may provide further insight into existing literature regarding TB and PB’s relative importance and refine suicide theory. The literature related to TB and PB’s relative importance as they relate to suicide ideation among psychiatric inpatients is mixed; some studies support TB or PB and others support TB and PB in relation to suicide ideation (for a review, see Mitchell et al., 2017). This may be due, in part, to the fact that facets (i.e., assessment items) of both TB and PB are likely related but also distinct from other facets within their respective constructs and suicide-related outcomes. For example, people may feel as though they belong more globally but may still endorse a lack of satisfying daily interactions or feel like an outsider in social gatherings, and, importantly, each of these facets may be differentially associated with passive and active suicide ideation. The use of item-level partial networks allows for the identification of how specific facets of TB and PB conditionally relate to each other and to passive or active suicide ideation, as well as which facets of TB and PB bridge the gap between TB and PB and passive or active suicide ideation. Furthermore, according to the IPTS, passive suicide ideation occurs when an individual experiences either TB or PB. However, in the context of other facets of TB and PB, it is possible that only certain facets of either TB or PB remain relevant to passive suicide ideation, which can be tested within an item-level network analysis approach. In addition, the IPTS would suggest that specific facets of TB and PB remain important and have unique associations with active suicide ideation.
An item-level network analysis approach may have important implications for assessment, prediction, and clinical intervention. A partial network allows for the identification of how specific facets of TB and PB conditionally relate to each other and to passive or active suicide ideation, which places focus on the unique clinical and predictive importance of specific facets. This facilitates the identification of the INQ items (i.e., facets) that are more strongly associated (or predictive in the context of causal models) with passive or active suicide ideation, which could be used to further refine the INQ and inform suicide risk assessment. Currently, the INQ includes multiple items that reflect TB and PB, which may be inadvertently reducing the predictive and clinical importance of TB and PB. Specifically, there may be items assessing TB and PB that, although moderately related to each other, are not uniquely associated with suicide ideation and may reduce the scale’s overall predictive ability. Understanding which INQ items are most strongly related to passive and active suicide ideation may also indicate that not all of the INQ items are necessary. If this were the case, the INQ could be tailored to specific populations, samples, and contexts by removing items that are less central to the network or that do not have direct relations with suicide-related outcomes. In addition, identification of key facets could guide clinical intervention such that targeting facets that have unique direct associations with suicide-related outcomes may be the most efficient pathway to reduce suicide risk. Likewise, targeting facets that emerge as highly central to TB and PB may most efficiently reduce symptoms throughout the overall network.
In sum, item-level network analysis may have important conceptual, assessment, predictive, and clinical implications for understanding suicide risk. In the current study, we aimed to demonstrate the utility of this approach by using an item-level network analysis approach to test the conditional relations predicted by the IPTS for the development of passive and active suicide ideation in a high-risk sample. First, we hypothesized that a few facets of TB or PB would positively and uniquely predict passive or active suicide ideation after controlling for all other facets within the network; however, given the lack of item-level research on the INQ, we do not have hypotheses about which specific facets. Theoretically, we would expect facets of both TB and PB to demonstrate unique associations with active suicide ideation, whereas relations between only TB or PB facets and passive suicide ideation may remain while controlling for all other facets in the network. Second, we hypothesized that facets of TB and PB would positively relate to each other and that specific facets would have stronger associations. Third, we hypothesized that facets of TB and PB would form separate, distinct clusters within the overall networks, which would suggest that these constructs could represent a set of highly related and reciprocally predictive facets. We also aimed to empirically identify facets that link different clusters of facets (e.g., “bridge symptoms”). As an exploratory aim, we tested the most central facets of TB and PB as indicated by the network analyses as predictors of passive and active suicide ideation compared with the full TB and PB subscales.
Method
Participants
Participants were 402 adult, English-fluent, psychiatric inpatients. The current study involved secondary data analysis and included portions of samples that have been analyzed and published previously (Cero et al., 2015; Jahn et al., 2015; Roush et al., 2018); however, before the current study, data from these samples have not been combined or analyzed in this manner. Individuals were excluded if they had a guardian advocate, were deemed by the research assistant to be unable to provide consent on the basis of their understanding of the information provided in the consent form, and/or were experiencing active psychosis or detoxification. As part of inclusionary criteria in Subsample 1, but not Subsamples 2 and 3, all participants were admitted to an inpatient unit for suicide risk concerns, and a summary of suicide risk based on participants’ responses was provided to the staff at the end of the research session. Subsample 1 (n = 103) was recruited from two psychiatric inpatient units in the southwestern United States. Subsample 2 (n = 118) was also recruited from a psychiatric inpatient unit in the southwestern United States. Subsample 2 data collection occurred at one of the same facilities as Subsample 1; data collection for Subsamples 1 and 2 did not overlap. Subsample 3 (n = 181) was collected from a psychiatric inpatient unit in the southeastern United States. For demographic data for the subsamples, see Table S1 in the Supplemental Material available online.
Procedures
All study procedures were conducted individually with participants during one assessment session. After providing consent, participants completed each assessment in a conference room or group room on the unit. Participants were permitted to discontinue participation at any time without penalty. Participants were not compensated for their participation. All procedures were approved by the appropriate university and hospital institutional review boards.
Measures
Interpersonal Needs Questionnaire
The INQ (Van Orden et al., 2012) is a 15-item self-report questionnaire that assesses TB (nine items) and PB (six items). Participants rate each item on a 7-point ordinal response scale from 1 (not at all true for me) to 7 (very true for me); higher scores indicate greater TB or PB. Several items were reverse-scored, so the direction of scoring was consistent across items. The INQ has demonstrated good convergent, discriminant, and construct validity (Van Orden et al., 2012). The INQ has also demonstrated good reliability in psychiatric inpatient samples (e.g., Cero et al., 2015; Monteith et al., 2013). In the current study, Cronbach’s αs for the INQ for the total sample were .86 and .93 for TB and PB, respectively. For this study, individual items, rather than the two subscales, were used as ordinal categorical facets in the network analyses. See Table S2 in the Supplemental Material for item-level frequencies and Table S3 in the Supplemental Material for item-level bivariate correlations with active and passive suicide ideation.
Beck Scale for Suicide Ideation
The Beck Scale for Suicide Ideation (BSS; Beck et al., 1997) is a 21-item self-report questionnaire assessing suicide ideation, intent, and plans over the preceding 4 weeks as well as previous suicide attempts. Participants rate each item on a 3-point ordinal response scale ranging from 0 to 2; higher scores indicate greater suicide ideation. The BSS has demonstrated good convergent validity with a clinician-rated assessment of suicide ideation in an inpatient sample (Beck et al., 1988). In this study, only two BSS items were used as facets: Item 2 (desire for death) and Item 4 (desire to kill oneself). As in previous research (Mitchell et al., 2017), we treated these items as ordinal categorical variables.
Data analysis approach
Following preliminary analyses, two regularized partial fixed effects matrices were estimated using the glmmLasso (Version 1.5.1; Groll & Tutz, 2014) for the R software environment (Version 3.6.1; R Core Team, 2020). Next, the matrices were graphed as networks, and centrality indices were estimated using qgraph (Version 1.4.4; Epskamp et al., 2012). Finally, the network’s community structure was examined using igraph (Version 1.1.2; Csardi & Nepusz, 2006).
Because these data are cross-sectional and observational in nature, causality cannot be inferred. However, the regularization techniques used led to a unique set of independent variables, which remain significant predictors in each dependent variable’s optimal regression model. Thus, the direction of the cross-sectional relation becomes important because it may be partialled on a different set of predictors than the inverse relation. Therefore, the directed networks and their associated centrality indices were used to demonstrate the unique strength and direction of prediction between nodes.
Networks
Before network development, the normality of each item’s response distribution was evaluated by examining frequencies, skewness, and kurtosis (Kline, 2005). Then, partial fixed effects were estimated using generalized linear mixed models in which each node was regressed on all other nodes simultaneously (Level 1), nested within the sampling site (Level 2). This allowed for the estimation of the fixed effects between nodes while controlling for within-site similarities. Because estimating such a large number of parameters can lead to spurious relations, random intercept models were estimated and regularized using the least absolute shrinkage and selection operator (LASSO; Friedman et al., 2008) with the R package glmmLasso (Groll & Tutz, 2014). The LASSO technique uses a tuning parameter to control the degree to which the model is penalized and unreliable parameters are shrunk to zero, thus resulting in a more parsimonious set of unique relations (Epskamp et al., 2018; van Borkulo et al., 2015). In this study, the Bayesian information criterion (BIC; Schwarz, 1978) was used to select each model’s optimal tuning parameter (Groll & Tutz, 2014). Finally, considering the ordinal nature of the BSS items, the IPTS assertion that active and passive suicide ideation develop differently, and literature that supports examining these outcomes separately (Mitchell et al., 2017, 2019), two separate regularized partial fixed effects matrices were developed in which either passive or active suicide ideation acted as the only ordinal dependent variable in the matrix; referred to hereafter as the passive and active networks, respectively.
In a single-level partial correlation network, edges represent the conditional associations between two nodes present in the used sample after controlling for other nodes in the network (Epskamp et al., 2012). In the current study, multiple sites were sampled, and the partial fixed effects were estimated using a multilevel partial correlation network in which edges represent the across-site average (i.e., population) relations between nodes that are over and above those explained by other nodes in the network. Furthermore, although temporal precedence is not established when using cross-sectional data, conditioning the estimated edges on the other facets in the network provides evidence that the predictive relation, demonstrated via the edge, is not better explained by other network nodes. Finally, the regularization technique uses a unique set of independent variables for each outcome, which provides evidence that the predictive relation direction, demonstrated via the arrow, is unique.
Centrality indices
Network figures do not provide a direct quantification of each node’s predictive ability (i.e., facet) on the entire network; thus, three centrality indices were estimated to aid in the interpretation of the network: outstrength, instrength, and betweenness. Within a cross-sectional network of partial relations, outstrength estimates how much variance a given node uniquely predicts in direct relations to other nodes, whereas instrength estimates how much of a given node’s variance is uniquely predicted by other nodes in the network. Betweenness counts the number of times a node lies on the shortest nondirect path between two other nodes (Bringmann et al., 2015). For each centrality index, higher values indicate greater centrality within the network (Opsahl et al., 2010).
Community structures
In complex networks, smaller clusters of nodes (i.e., communities) may be identified within the larger network. For the present study, the spinglass function (Reichardt & Bornholdt, 2006) was used to assess for separate communities within the glasso network. This function aims to find communities of nodes that have a high number of positive edges among themselves but fewer positive edges outside of the community. This is reversed for negative relations (i.e., there are fewer negative relations within the community compared with outside the community). In other words, communities consist of a smaller cluster of nodes that relate more strongly to each other than to nodes outside the community. For further detail, see Reichardt and Bornholdt (2006).
Logistic regression
Ordinal logistic regression analyses were conducted in SAS (Version 9.4). Given the network results, the single items of TB and PB that emerged as more central to the networks were examined as predictors of passive and active suicide ideation and then compared with logistic regression analyses examining TB and PB subscale scores as predictors to demonstrate their relative predictive ability. We also tested the statistical significance of the difference between logit estimates (see Paternoster et al., 1998).
Results
Figures 1 and 2 depict the inferred passive and active networks, respectively. The strength of these predictive relations is visually depicted by the thickness of the edge (e.g., line in the figure), and the direction is indicated by the arrow.
Passive network. The stronger the predictive relation, the thicker the line. The direction of the relation is indicated by the arrow. Passive = Beck Scale for Suicide Ideation Item 2, desire for death; Better Off Without Me = Interpersonal Needs Questionnaire (INQ) Item 1, belief that people “would be better off if I were gone”; Happier Without Me = INQ Item 2, belief that people “would be happier without me”; Burden on Society = INQ Item 3, feeling like “a burden on society”; My Death a Relief to Others = INQ Item 4, belief that “my death would be a relief” to others; Could Be Rid of Me = INQ Item 5, belief that people “wish they could be rid of me”; I Make Things Worse = INQ Item 6, belief that “I make things worse for the people in my life”; Others Care About Me = INQ Item 7, belief that others “care about me”; Feel Like I Belong = INQ Item 8, feeling “like I belong”; Interact With Caring Others = INQ Item 9, belief that “I rarely interact with people who care about me”; Have Many Friends = INQ Item 10, belief that “I am fortunate to have many caring and supportive friends”; Disconnected = INQ Item 11, feeling “disconnected from other people”; Outsider = INQ Item 12, feeling like an “outsider in social gatherings”; Times of Need = INQ Item 13, feeling that there are “people I can turn to in times of need”; Close to Others = INQ Item 14, feeling “close to other people”; Satisfying Interaction = INQ Item 15, belief about having “at least one satisfying interaction” daily.
In Figure 1, the strongest connection between passive suicide ideation and TB is the predictive relation from the facet of “feeling that one belongs” to passive suicide ideation (partial r = .41, p = .002). Furthermore, the strongest connection between passive suicide ideation and PB is the predictive relation from the facet of “feeling like a burden on society” to passive suicide ideation (partial r = .63, p < .001). In Figure 2, similar to Figure 1, the strongest connection between active suicide ideation and TB is the relation from the facet of “feeling that one belongs” to active suicide ideation (partial r = .59, p < .001). Furthermore, the strongest connection between active suicide ideation and PB is the relation from the facet of “feeling like a burden on society” to active suicide ideation (partial r = .45, p = .013).
Active network. The stronger the predictive relation, the thicker the line. The direction of the relation is indicated by the arrow. Active = Beck Scale for Suicide Ideation Item 4, desire to kill oneself; Better Off Without Me = Interpersonal Needs Questionnaire (INQ) Item 1, belief that people “would be better off if I were gone”; Happier Without Me = INQ Item 2, belief that people “would be happier without me”; Burden on Society = INQ Item 3, feeling like “a burden on society”; My Death a Relief to Others = INQ Item 4, belief that “my death would be a relief” to others; Could Be Rid of Me = INQ Item 5, belief that people “wish they could be rid of me”; I Make Things Worse = INQ Item 6, belief that “I make things worse for the people in my life”; Others Care About Me = INQ Item 7, belief that others “care about me”; Feel Like I Belong = INQ Item 8, feeling “like I belong”; Interact With Caring Others = INQ Item 9, belief that “I rarely interact with people who care about me”; Have Many Friends = INQ Item 10, belief that “I am fortunate to have many caring and supportive friends”; Disconnected = INQ Item 11, feeling “disconnected from other people”; Outsider = INQ Item 12, feeling like an “outsider in social gatherings”; Times of Need = INQ Item 13, feeling that there are “people I can turn to in times of need”; Close to Others = INQ Item 14, feeling “close to other people”; Satisfying Interaction = INQ Item 15, belief about having “at least one satisfying interaction” daily.
Taken together, these findings suggest that as hypothesized, a few key facets of TB and PB uniquely predict passive and active suicide ideation after controlling for all other facets within the network. Furthermore, as hypothesized, key facets from both TB and PB remained central and uniquely predictive of active suicide ideation. Specifically, the responses to the direct questions regarding lower feelings of belonging and greater feelings of being a burden on society are most strongly uniquely predictive of both passive and active suicide ideation. Face-valid items with wording that directly match the latent constructs of TB and PB proposed by the IPTS may be particularly important in predicting and assessing both passive and active suicide ideation.
Centrality indices
Figures 3 and 4 display the standardized results of the centrality analyses for the passive and active networks, respectively. In Figure 3, outstrength estimates indicate facets of “feelings of being a burden on society” (1.70) and thinking people “would be happier without you” (1.23) were the most uniquely predictive of other nodes in the passive network. Instrength estimates indicate that facets of passive suicide ideation (2.74) and the “feeling that one belongs” (1.79) were most strongly uniquely predicted by other nodes in the passive network. Finally, betweenness estimates suggest that facets of “feeling that one belongs” (2.60) and “feelings of being a burden on society” (1.45) were important for connecting nodes throughout the network and appear to have the most “mediating” connections within the passive network. This suggests these facets tend to be more proximal to a greater number of other facets.
Centrality indices for passive network. In the outstrength plot, the more positive the value of the node, the more predictive it was of other nodes in the passive network. In the instrength plot, the more positive the value of the node, the more predicted it was by the other nodes in the passive network. In the betweenness plot, the more positive the value of the node, the more important for connecting nodes throughout the network. Passive = Beck Scale for Suicide Ideation Item 2, desire for death; Better Off Without Me = Interpersonal Needs Questionnaire (INQ) Item 1, belief that people “would be better off if I were gone”; Happier Without Me = INQ Item 2, belief that people “would be happier without me”; Burden on Society = INQ Item 3, feeling like “a burden on society”; My Death a Relief to Others = INQ Item 4, belief that “my death would be a relief” to others; Could Be Rid of Me = INQ Item 5, belief that people “wish they could be rid of me”; I Make Things Worse = INQ Item 6, belief that “I make things worse for the people in my life”; Others Care About Me = INQ Item 7, belief that others “care about me”; Feel Like I Belong = INQ Item 8, feeling “like I belong”; Interact With Caring Others = INQ Item 9, belief that “I rarely interact with people who care about me”; Have Many Friends = INQ Item 10, belief that “I am fortunate to have many caring and supportive friends”; Disconnected = INQ Item 11, feeling “disconnected from other people”; Outsider = INQ Item 12, feeling like an “outsider in social gatherings”; Times of Need = INQ Item 13, feeling that there are “people I can turn to in times of need”; Close to Others = INQ Item 14, feeling “close to other people”; Satisfying Interaction = INQ Item 15, belief about having “at least one satisfying interaction” daily.
In Figure 4, outstrength estimates indicate facets of “feelings of being a burden on society” (1.61) and thinking that “one has many caring and supportive friends” (1.22) were the most uniquely predictive of other nodes in the active network. Instrength estimates indicate that facets of active suicide ideation (1.81) and thinking that “one’s death would be a relief to others” (0.93) were most strongly uniquely predicted by other nodes in the active network. Finally, betweenness estimates suggest that facets of “feeling like a burden on society” (2.18) and thinking that “one’s death would be a relief to others” (1.70) were important for connecting nodes throughout the network and appear to have the most mediating connections within the active network.
Centrality indices for active network. In the outstrength plot, the more positive the value of the node, the more predictive it was of other nodes in the passive network. In the instrength plot, the more positive the value of the node, the more predicted it was by the other nodes in the passive network. In the betweenness plot, the more positive the value of the node, the more important for connecting nodes throughout the network. Active = Beck Scale for Suicide Ideation Item 4, desire to kill oneself; Better Off Without Me = Interpersonal Needs Questionnaire (INQ) Item 1, belief that people “would be better off if I were gone”; Happier Without Me = INQ Item 2, belief that people “would be happier without me”; Burden on Society = INQ Item 3, feeling like “a burden on society”; My Death a Relief to Others = INQ Item 4, belief that “my death would be a relief” to others; Could Be Rid of Me = INQ Item 5, belief that people “wish they could be rid of me”; I Make Things Worse = INQ Item 6, belief that “I make things worse for the people in my life”; Others Care About Me = INQ Item 7, belief that others “care about me”; Feel Like I Belong = INQ Item 8, feeling “like I belong”; Interact With Caring Others = INQ Item 9, belief that “I rarely interact with people who care about me”; Have Many Friends = INQ Item 10, belief that “I am fortunate to have many caring and supportive friends”; Disconnected = INQ Item 11, feeling “disconnected from other people”; Outsider = INQ Item 12, feeling like an “outsider in social gatherings”; Times of Need = INQ Item 13, feeling that there are “people I can turn to in times of need”; Close to Others = INQ Item 14, feeling “close to other people”; Satisfying Interaction = INQ Item 15, belief about having “at least one satisfying interaction” daily.
Community structures
The spinglass method (Reichardt & Bornholdt, 2006) was used to assess for distinct clusters within the larger passive and active networks. This method showed that the modularity estimates for both the passive (−0.03) and active (−0.29) networks were negative, indicating that the degree of clustering was less than would be expected using random data and is likely spurious. Therefore, contrary to hypotheses, TB and PB facets do not form distinct communities within the overall network but rather appear intermingled. As a result, “bridge symptoms” were unable to be identified.
Logistic regression
Given the network results, the face-valid items reflecting feelings of belonging and being a burden on society emerged as most uniquely predictive of both active and passive suicide ideation; thus, these items were examined as predictors of passive and active ideation compared with the full TB and PB subscales in logistic regression analyses.
The model with the single TB and PB items as predictors of passive suicide ideation fit the data significantly better (Akaike information criterion [AIC] = 668.73, Nagelkerke R2 = .36) than a model with the full TB and PB subscales as predictors (AIC = 685.44, Nagelkerke R2 = .32). In the single-items model, feeling like one belongs (odds ratio [OR] = 1.26, d = 0.13, p < .001) and feeling like a burden on society (OR = 1.53, d = 0.24, p < .001) were associated with greater odds of elevated passive suicide ideation. In the subscale model, the TB subscale (OR = 1.45, d = 0.21, p < .001) and the PB subscale (OR = 1.52, d = 0.23, p < .001) were associated with greater odds of elevated passive suicide ideation. The logit estimates for the feeling that one belongs item (logit = 0.23, SE = 0.06) and the TB subscale (logit = 0.37, SE = 0.09) predicting passive suicide ideation were not statistically significantly different (z = −1.33, p = .184, two-tailed), which suggests that the additional items in the TB subscale do not improve predictive ability. The logit estimates for the feeling like a burden on society item (logit = 0.43, SE = 0.06) and PB subscale (logit = 0.42, SE = 0.07) predicting passive suicide ideation were also not statistically significantly different (z = 0.10, p = .920, two-tailed).
The model with the single TB and PB items as predictors of active suicide ideation fit the data significantly better (AIC = 659.30, Nagelkerke R2 = .30) than the model with the full TB and PB subscales as predictors (AIC = 665.58, Nagelkerke R2 = .29). In the single-items model, feeling like one belongs (OR = 1.33, d = 0.16, p < .001) and feeling like a burden on society (OR = 1.42, d = 0.19, p < .001) were associated with greater odds of elevated active suicide ideation. In the subscale model, the TB subscale (OR = 1.54, d = 0.24, p < .001) and the PB subscale (OR = 1.45, d = 0.21, p < .001) were associated with greater odds of elevated active suicide ideation. The logit estimates for the feeling that one belongs item (logit = 0.28, SE = 0.06) and the TB subscale (logit = 0.43, SE = 0.10) predicting active suicide ideation were not statistically significantly different (z = 1.29, p = .197, two-tailed), which suggests that the additional items in the TB subscale do not improve predictive ability. The logit estimates for the feeling like a burden on society item (logit = 0.35, SE = 0.06) and PB subscale (logit = 0.37, SE = 0.07) predicting active suicide ideation were also not statistically significantly different (z = 0.17, p = .865, two-tailed).
Discussion
Network analysis has only recently been used to examine risk for suicidal behaviors (de Beurs et al., 2017, 2019), and this initial work provides support for conceptualizing suicide risk as a complex system of interconnected risk factors that have unique and meaningful associations with relevant suicide-related outcomes. However, broader literature, including these studies, has tested latent factors; this limits the understanding of suicide risk factors to broader constructs, which may also constrain the ability to accurately assess, predict, and effectively manage suicide risk.
In the current study, we aimed to provide support for the broader utility of using an item-level network analysis approach to provide a more fine-grained conceptualization of suicide risk. To demonstrate the conceptual, assessment, predictive, and clinical implications of this approach, we used an item-level network analysis to test the conditional relations predicted by the IPTS for the development of passive and active suicide ideation in a sample of psychiatric inpatients. The current study provides insight into whether the relations among TB, PB, and passive and active suicide ideation are equivalently influential across all facets or whether the relations among these constructs are driven by a few key facets that may be more important in assessing and managing suicide risk.
The current findings have direct implications for the conceptualization of suicide risk and suicide theory. Contrary to IPTS theory, the community analyses results indicate that facets of TB and PB do not form distinct clusters, but instead, they are intermingled throughout the overall network. Furthermore, as hypothesized, the current study results suggest that after partialling out the other facets in the network, only a few facets may be influential for both active and passive suicide ideation. That is, not all the facets of TB and PB demonstrated predictive validity of passive and active suicide ideation. Rather, there were a select few strong connections exhibited among facets of TB, PB, and passive and active suicide ideation. Specifically, feeling that one belongs and feeling like a burden on society were the only two facets uniquely predictive of both passive and active suicide ideation.
Furthermore, the idea that a select few facets from each construct are playing a key role within the overall networks was supported by the centrality indices. For example, in terms of outstrength, feelings of being a burden on society was the most uniquely predictive facet of other nodes in both the passive and active networks. Note that some facets emerged as more uniquely predictive in the context of passive suicide ideation compared with active suicide ideation. Specifically, thinking people would be “happier without you” seems to be uniquely predictive of passive suicide ideation but not active suicide ideation. Similar results were found for instrength and betweenness; a few facets appeared to be more uniquely predicted by other nodes and had more indirect connections in the network, respectively. The results from the logistic regression analyses provide additional support for the relative importance of key facets such that the individual face-valid items identified in the network results fit the data better than the subscale score models and the subscale scores were not significantly better predictors of active and passive suicide ideation compared with the single items. Taken together, TB and PB may be better conceptualized as an interconnected constellation of thwarted interpersonal needs (i.e., TB and PB) that have unique associations with suicide-related outcomes, with only a few key facets that are central and uniquely predictive.
The fact that the relations between thwarted interpersonal needs and passive and active suicide ideation were driven by a few key facets within each construct has important implications for how researchers assess these two constructs of the IPTS moving forward and may help explain the discordant results of the current IPTS literature. Much of the research examining the role of TB and PB as measured by the INQ (Van Orden et al., 2012) is based on analyses that use TB and PB subscale scores as predictors of suicide-related outcomes, including the theory-based network analysis of suicide risk conducted by de Beurs et al. (2019). The limitation of this approach is that even if someone scores highly on the key facets of TB and PB, the relations between these facets and passive and active suicide ideation may appear to be less significant because that facet’s predictive ability or relative importance may be diluted or weakened by low scores on less relevant facets. For example, our results suggested that feeling like a burden on society might drive the relation between PB and both passive and active suicide ideation. Thus, if someone scored highly on this item (i.e., facet), we might expect him or her to endorse a greater desire for death or suicide; however, if this person also scored low on the other facets of PB, such as believing that people wish they could be rid of him or her, then the subscale score would appear to be relatively average. This would directly affect the ability to accurately assess and predict this individual’s suicide risk level such that expectations for the individual’s risk for passive or active suicide ideation would be diminished when in fact, a particular facet may more accurately identify the individual at higher risk for suicide ideation. This is especially true if, as our data suggest, there are only one or two key facets within each construct and the ratio of facets that are directly predictive of suicide ideation compared with those that are only indirectly related is low. These findings suggest that TB and PB may be most accurately assessed by a limited number of items, which could improve the predictive validity of the IPTS.
From a broader perspective, the current findings demonstrate the utility of an item-level network analysis approach for testing and refining complex suicide risk models. A more nuanced perspective of the role of suicide risk factors in the context of a myriad of other risk factors allows for identifying key risk factors that can be used to improve assessment and prediction of suicide risk. The use of these key predictive factors or items, those that emerge as the most central and predictive within a network, could also be used to refine other methodological approaches to the study of suicide risk. For example, ecological momentary assessment (EMA) necessitates using a limited number of items, yet there is no consistent data-driven approach for item selection. Item selection for EMA could be optimized to include key items that emerge as more predictive of relevant outcomes within a suicide risk network.
There are several noteworthy implications of an item-level network analysis approach that may serve to bridge the gap between research and clinical practice. Within an item-level network analysis approach, there is a greater emphasis on actual facets or nodes, which clarifies the importance of specific facets that potentially may cause or maintain feelings of TB, PB, and passive or active suicide ideation. Future studies may include risk or protective factors for suicide such as TB and PB as well as other theory-based variables (e.g., hopelessness, instrumental social support, serotonergic dysfunction) and can be modeled in one network, allowing for the identification of specific facets that might influence other facets in the suicide ideation network. These specific facets may, in turn, account for the relation between suicide ideation and other risk factors, thus explaining the entire network through a particular path or sequence of paths. Identifying and targeting these key facets would be the most effective and efficient treatment, such that targeting one key facet would result in the greatest reductions in suicide risk because of its important relations with other risk factors. For example, the betweenness estimates of this study suggest that feeling that one belongs has the most mediating connections in the passive suicide ideation network, whereas feelings of being a burden on society had the most mediating connections in the active suicide ideation network. The findings of this study suggested face-valid items with wording that directly matches the latent constructs of TB and PB as proposed by the IPTS (i.e., feelings of being a burden on society, feeling that one does not belong) may drive the relation with passive and active suicide ideation; thus, targeting these facets may be the most effective and efficient means of reducing suicide risk.
Identifying individually predictive facets would also be important in assessing suicide risk across different contexts, such as in community mental health and acute psychiatric inpatient settings where patients are being assessed for a wide array of symptoms and rapid identification of suicide risk level is needed. In such settings, the assessment of each construct is often limited to a single item. Thus, understanding which suicide risk factors or, in this case, which facets of TB and PB are driving the relation with passive and active suicide ideation would allow clinicians in these types of settings to develop the most efficient and accurate suicide risk assessments. Pending replication, our results would suggest that directly assessing whether people feel like a burden on society or feel as if they belong may provide the most insight into risk for both passive and active suicide ideation.
This study served as the initial step toward using an item-level network analysis approach to better examine and refine complex suicide risk models. For example, we sought to identify key facets of the IPTS that predict passive and active suicide ideation and provided initial evidence for the important conceptual, assessment, predictive, and clinical implications of an item-level approach. We used psychiatric inpatient samples, which improves the understanding of suicide risk in a high-risk group; however, these findings may not generalize to other more demographically and psychiatrically diverse samples. In addition, although the directed networks indicated the predictive direction of relations, causality cannot be inferred because of the cross-sectional nature of the data. Finally, although we examined the role of TB and PB, which are posited to be the most proximal risk factors to passive and active suicide ideation according to the IPTS (Joiner, 2005; Van Orden et al., 2010), we were unable to examine other risk factors (e.g., depression, hopelessness) in the current networks because of concerns about the number of parameters being estimated with the relatively small sample size.
Future studies should attempt to replicate the current findings in various samples to determine whether a few key facets emerge as robust predictors of passive and active suicide ideation across samples or whether key facets are sample-specific. Future studies with larger samples may also consider examining more facets of suicide ideation (e.g., frequency, duration, intent) to provide a more thorough understanding of the relations among facets of TB, PB, and suicide ideation. Although the current study did not find evidence for distinct communities or clusters, future researchers should consider the potential for items with similar wording or factors obtained with similar methods to form separate clusters. Future studies may also consider examining the relations between facets of other theory-based risk factors (e.g., hopelessness, suicide capability) and suicide risk. Future longitudinal studies using this method would allow for the determination of directional (i.e., dynamic, temporal, or potentially causal) relations between facets, which might allow for the identification of specific empirically supported treatment targets that can be included within an evidence-based psychological treatment (e.g., cognitive behavior therapy). If specific facets that are more central to the network are targeted, rather than attempting to reduce an entire construct or cluster of facets simultaneously, beneficial effects may propagate through the network across facet clusters. A network analysis approach may also be used to develop person-specific networks (David et al., 2018) such that individuals may have different networks that would suggest targeting different facets of TB and PB to reduce passive and active suicide ideation. Repeated assessment of specific facets throughout treatment may allow clinicians to monitor and adjust treatment using these person-specific networks (Robinaugh et al., 2014).
The current study was the first to our knowledge to use and demonstrate the utility of an item-level network analysis to the study of suicide risk. Specifically, we examined the hypothesized relations of the IPTS in a high-risk sample of psychiatric inpatients using a multilevel approach to identify key facets associated with passive and active suicide ideation. Results suggest a theory-driven, item-level network analysis approach may provide a more detailed or fine-grained conceptualization of suicide risk given the emphasis on the meaningful associations between individual risk factors (i.e., facets, items) and relevant suicide-related outcomes. Future use of an item-level network analysis approach for the study of suicide may address important gaps in the literature, such as dynamic relations between facets of theory-based constructs, variation across samples, and relative importance of facets that can further inform risk assessment and treatment.
Supplemental Material
sj-pdf-1-cpx-10.1177_21677026211000670 – Supplemental material for Suicide Ideation and Thwarted Interpersonal Needs Among Psychiatric Inpatients: A Network Approach
Supplemental material, sj-pdf-1-cpx-10.1177_21677026211000670 for Suicide Ideation and Thwarted Interpersonal Needs Among Psychiatric Inpatients: A Network Approach by Sarah L. Brown, Andrew J. Marshall, Sean M. Mitchell, Jared F. Roush, Gregory H. Mumma, Danielle R. Jahn, Jessica D. Ribeiro, Thomas E. Joiner and Kelly C. Cukrowicz in Clinical Psychological Science
Footnotes
Transparency
Action Editor: Christopher G. Beevers
Editor: Scott O. Lilienfeld
Author Contributions
S. L. Brown, A. J. Marshall, J. F. Roush, and S. M. Mitchell developed the study concept and drafted the manuscript. A. J. Marshall and S. L. Brown performed the data analysis. D. R. Jahn, J. D. Ribeiro, J. F. Roush, S. L. Brown, and S. M. Mitchell performed data collection. All of the authors provided critical revisions and approved the final manuscript for submission.
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
