Time Varying Prediction of Thoughts of Death and Suicidal Ideation in Adolescents: Weekly Ratings over 6-month Follow-Up

Abstract

Suicidal ideation (SI) and thoughts of death are often experienced as fluctuating; therefore a dynamic representation of this highly important indicator of suicide risk is warranted. Theoretical accounts have suggested that affective, behavioral, and interpersonal factors may influence the experience of thoughts of death/SI. This study aimed to examine the prospective and dynamic impact of these constructs in relation to thoughts of death and SI. We assessed adolescents with a recent hospitalization for elevated suicide risk over 6 months. Using the methodology of the Longitudinal Interval Follow-Up Evaluation, weekly ratings for SI, course of depressive illness, affect sensitivity, negative affect intensity, behavioral dysregulation, peer invalidation, and family invalidation were obtained. Using multilevel modeling, results indicated that (a) same-week ratings between these constructs and SI were highly correlated at baseline and throughout follow-up; (b) baseline ratings of affect sensitivity, behavioral dysregulation, and peer invalidation were positive prospective predictors of SI at any week of follow-up; (c) weekly ratings of each of these constructs had significant associations with next-week ratings of SI; and (d) ratings of SI had positive significant associations with next-week ratings on each of the constructs. These results suggest that affective sensitivity, behavioral dysregulation, peer invalidation, and SI are highly associated with SI levels both chronically (over months) and acutely (one week to the next), whereas depression, negative affect intensity, and family invalidation were more acutely predictive of SI. Elevated SI may then aggravate all these factors in a reciprocal manner.

Acknowledgments

Funding for this project was provided by National Institute of Mental Health grant K23 MH069904 to Shirley Yen.

Notes

Note: CSR = Construct Status Rating.

¹These κ values are higher than κ values for individual disorders likely because CSR variables consisted of only one scale score versus multiple scores for each of many diagnostic criteria for a diagnosis.

Note: N = 121. SI = thoughts of death and suicidal ideation; DEP = depression; AS = affective sensitivity; NAI = negative affective intensity; BD = behavioral dysregulation; PeerI = peer invalidation; FamI = family invalidation.

*p < .05. **p < .001.

Note: N = 99. SI = thoughts of death and suicidal ideation; DEP = depression; AS = affective sensitivity; NAI = negative affective intensity; BD = behavioral dysregulation; PeerI = peer invalidation; FamI = family invalidation.

**p < .001.

²Regarding the HGLM models, the response distribution for SI was Poisson, which accounts for the count-nature distribution of the SI-CSR variable, with a natural logarithm transformation link so that it is consistent with the Poisson distribution. Level 1 assesses the weekly predictors of SI by adjusting the individual Level 2 intercept, and Level 2 assesses the baseline predictors of SI and includes a random intercept, with individual level error in measurement of weekly SI. The base model was an individual-level intercept for weekly SI modified by the week plus random error. The predictor model involved adding specific predictors to the base model. From each of the predictor weights we also calculated RRs, which are standardized indices that indicate amount of changes in the level of the outcome variable (SI) relative to the change of each predictor variable by one standardized unit. Model comparison was evaluated with the following fit-indices: log-likelihood (H ₀ ), Akaike information criterion (AIC), and Bayesian information criterion (BIC). Larger H ₀ values indicate better fit, whereas smaller AIC and BIC values indicate better fit (Burnham & Anderson, 2004). Based on these fit indices, the model with predictors (H ₀  = −4,769.61, AIC = 9,569.21, BIC = 9,655.19) provided incremental fit to the data beyond the baseline model (H ₀  = −34,218.70, AIC = 68,441.40, BIC = 68,453.20). There was a significant random intercept in both the base model (β = 2.31, SE = .011, p < .001) and the predictor model (β = 5.64, SE = .001, p < .001).

³Upon initial analysis, the majority of the week-prior predictors were significant inverse predictors of SI, contrary to our hypotheses and previous research. This unexpected finding indicated that suppression effects might have resulted in the sign reversals for some predictors. Based on the recommendations by Gaylord-Harden, Cunningham, Holmbeck, and Grant (2010), one major potential for net suppression (where the direction of a variable is the opposite direction as theoretically predicted) was the inclusion of multiple correlated measures in the prediction of SI. The recommended way to test for such suppression effects is to run the analysis with all predictors to determine significance but then to run each predictor individually to determine if there was a change in the direction of the relationship. Gaylord-Harden et al. recommend that if the univariate analysis is in the expected direction, then the negative direction of the multivariate weight should not be interpreted as being in the opposite direction hypothesized.

Upon examination with this approach, we indeed discovered that suppression was accounting for the reverse direction of most of the variables. As we originally expected, and consistent with previous research, week-prior affective sensitivity, negative affect intensity, peer invalidation, and depression all significantly and positively predicted subsequent week SI level when examined individually. As in the original model, week-prior family invalidation, behavioral dysregulation, and SI level maintained significant positive associations with SI. We also examined the significant baseline predictors in the model to ensure that suppression effects were not occurring with these variables. In these follow-up analyses, however, suppression effects were not indicated as all of the significant baseline predictors from the model maintained their significant paths in the original direction.