Rumination Mediates the Relation of Hostile Attribution to Psychological Maladjustment Among Adolescents from Three Countries

Participants were mothers, fathers, and children from three countries—Colombia, Italy, and the U.S.—enrolled in the Parenting Across Cultures Study (e.g., Lansford et al., 2014). Data collection occurred in four geographic regions: Medellín, Colombia; Naples and Rome, Italy; and Durham, North Carolina, U.S. Within the U.S., we analyzed data from three racial/ethnic groups: White, Black, and Latinx. We sought samples from countries/regions varying across a range of socio-demographic characteristics, including race, ethnicity, religion, parental education levels/socioeconomic status, and child well-being indicators. Our three countries varied on their Health Development Index score (HDI; i.e., a composite indicator that ranks countries based on health, education, and income). For example, based on 2012 HDI data (aligned with the middle time points in our data), the U.S. ranked 5th, Italy 28th, and Colombia 98th out of 187 countries. We also sought samples within the U.S. that varied on socio-historical and cultural experiences (e.g., differences in exposure to environmental stressors, experiences of discrimination) which could influence our variables of interest². Despite being quite diverse, participants came from convenience samples and are therefore not representative of entire cultures, nations, regions, or racial/ethnic groups.

Our total sample included 532 mothers, 384 fathers, and 566 children (siblings excluded; 50% girls). Sample sizes across subgroups were: Medellín = 100 adolescents, 85 mothers, 74 fathers; Naples = 95 adolescents, 94 mothers, 73 fathers; Rome = 99 adolescents, 99 mothers, 77 fathers; U.S. White = 100 adolescents, 95 mothers, 67 fathers; U.S. Black = 92 adolescents, 87 mothers, 50 fathers; and U.S. Latinx = 80 adolescents, 72 mothers, 43 fathers. Detailed socio-demographic information about families within each subgroup is presented in Supplemental Materials Table S1and in Iselin et al. (2022). In short, there was variability across the six subgroups in marital status (e.g., 37–83% married parents), parent education level (e.g., 10 to 16 years, on average), and percent working mothers (42–78%).

We analyzed data gathered at four time points [T1: youth Mage = 10.89 years (SD = 0.70); T2: youth Mage = 12.58 years (SD = 0.68); T3: youth Mage = 13.71 years (SD = 0.67); T4: youth Mage = 14.99 years (SD = 0.73)]. Youth-reported HAB was assessed at T1 and T3, youth-reported HR at T2, and parent-reported aggression, anxiety, and depression symptoms at T1, T2, T3, and T4. The overall retention rate was over 95% within the full sample across all time points. T1, T2, T3, and T4 data were gathered in 2010–2011, 2012–2013, 2012–2014, and 2014–2016, respectively.

We recruited participants from a range of public and private schools serving low- to high-income families by sending letters about the study home with students for their parents to read, sign, and return if they were interested in participating. We contacted interested parents via phone to schedule interviews at a location of their choosing.

We used a rigorous procedure of independent translation to ensure linguistic and conceptual equivalence across English, Spanish, and Italian versions of our measures (Erkut, 2010). Questionnaires were forward- and back-translated by graduate students, faculty, or staff members at the collaborating institutions who were fluent in English and the target language (Spanish or Italian). Translators were asked to (1) note where items did not translate well or were culturally insensitive/inappropriate, (2) identify words that elicited multiple meanings, (3) offer suggestions for improvement for identified problems, and (4) indicate reasons for suggesting changes. Site principal investigators (who were native to the country in which data was being collected and were faculty members at collaborating universities) met with the translators to review item-level discrepancies and ambiguities and modify items as needed. Our translation methods were designed to ensure measures were valid in all sites regarding linguistic equivalence as well as cultural meanings imparted by the measures (Erkut, 2010; Peña, 2007).

Before completing questionnaires, parents provided informed consent and youth provided informed assent to participant. Parent and youth participants completed questionnaires either on their own or with interviewers reading questions aloud and recording their answers. Interviewers were trained graduate students or staff members at the collaborating universities. Parent surveys took approximately 30 min to complete, and youth surveys took approximately 1½ hours to complete. We provided modest financial compensation in amounts deemed acceptable by Institution Review Boards (IRB) at each university. IRBs within each region/country approved all testing procedures and protocols (Duke University, Sapienza University of Rome, Federico II Second University of Naples, and Universidad San Buenaventura).

Our aim was to examine whether HR mediated the relation of HAB to aggression, anxiety, and depression symptoms, as well as whether HAB mediated the relation of HR to aggression, anxiety, and depression symptoms. HR was assessed at only a single timepoint (T2), so the timing varied between the two types of mediation models. Specifically, models testing HR as a mediator assessed HAB at T1, HR at T2, and outcomes at T3, while models testing HAB as a mediator assessed HR at T2, HAB at T3, and outcomes at T4.

Mediation model coefficients were obtained using Mplus 8.7 with full information maximum likelihood (FIML) estimation. FIML accounts for missing data by using all available data from each case to estimate parameters and adjust for potential bias in the estimates resulting from missing data. This approach provides unbiased estimates in the presence of data missing at random and has been identified as one of the optimal ways to handle missing data (Peugh & Enders, 2004). Mediation analyses controlled for youth baseline age (to control for developmental differences as a possible confounding variable), youth gender (to control for differences between boys and girls that may occur on our predictors and outcomes), and baseline levels of the outcome as covariates. We do not control for baseline levels of the mediator to ensure that all structural paths were parallel between our two models (see below). Indirect effects were tested using bootstrap confidence intervals (based on 5,000 resamples).

We used structural equation modeling (SEM) in our analyses to improve the precision of our estimates by modelling the measurement errors and unexplained variances (Hoyle & Smith, 1994). However, our sample size was on the lower end of what would be considered acceptable for SEM (O’Rourke & Hatcher, 2013). We therefore decided to combine scale items into parcels prior to performing our analyses to reduce the number of indicators in our models, stabilize our parameter estimates, and make our models more likely to converge (Matsunaga, 2008). This approach is particularly beneficial in multigroup analyses, where model complexity increases as additional groups are included in the comparisons. Parceling reduces the number of parameters that need to be estimated across groups, making the models more manageable, especially in the context of smaller sample sizes (Little et al., 2013).

For each latent construct (HAB, HR, aggression, anxiety, and depression), we created three domain-representative parcels by averaging conceptually similar items (Kishton & Widaman, 1994). Intentionally designed parcels are typically preferred because they provide meaning to the parcels and help reduce problematic sources of variance like correlated residuals or cross-loadings, which might otherwise confound the measurement model (Little et al., 2022). Intentional parceling commonly leads to more accurate parameter estimates and enhanced predictive validity. Parcels for HAB pertained to three different peer provocation contexts: personal property being ruined (3 items), physical discomfort (4 items), and schoolwork being wrecked (3 items). Parcels for HR pertained to three domains: sustaining/escalating hostility (3 items), remembering insults (2 items), and not forgiving insults (2 items). Parcels for aggression pertained to defiance (6 items), physical aggression (6 items), and negativity (6 items). Parcels for anxiety pertained to social doubts (2 items), perfectionism (2 items), and general apprehension (2 items). Parcels for depression pertained to isolation (2 items), low mood (2 items), and low self-worth (2 items). For measures of psychological adjustment (aggression, anxiety, and depressive symptoms), we defined the parcels the same way for both the baseline and outcome assessments. This consistency is critical for testing measurement invariance, a necessary step in multigroup analyses (Little et al., 2022). To measure model fit, we considered the χ² statistic, RMSEA (Root Mean Square Error of Approximation), CFI (Comparative Fit Index), and SRMR (Standardized Root Mean Square Residual) (see Kenny, 2015). Better models are associated with higher CFI and lower χ², RMSEA, and SRMR values.

We were interested in examining whether our mediation models applied similarly across our national, regional, and racial subgroups, as well as across genders. Given that the diversity within our sample is a critical source of variability in our variables of interest, we believe that comparing model fit across subgroups provides a unique and robust examination of whether meditational pathways function similarly among diverse youth. The multigroup analyses tested the two coefficients composing the mediation effect (the relation of the predictor to the mediator and the relation of the mediator to either aggression, depression symptoms, or anxiety symptoms after controlling for the predictor) as well as the direct effect of the predictor on either aggression, depression symptoms, or anxiety symptoms independent of the mediated effect. The multigroup analyses tested whether a model allowing these coefficients to be estimated independently within each subgroup fit the data significantly better than a model where these coefficients were constrained to be the same across subgroups. If constraining the coefficients to be equal across subgroups did not significantly worsen model fit, we concluded our mediation effects were consistent across subgroups. Coefficients for the covariates (i.e., baseline symptoms and age) were allowed to freely vary among groups in all models.

Table 1 presents descriptive statistics for study variables. For all variables of interest, mean scores were below each scale’s mid-point, suggesting mild severity or frequency of HAB, HR, aggression, depression, and anxiety. Graphical and statistical examinations of normality (i.e., univariate skewness < 2.5, kurtosis < 7.0) did not suggest any substantial deviations from normality for numeric variables (see Supplemental Materials Table S2). Table 1 also presents overall reliability estimates for all composites (reliability estimates within each subsample are presented in Supplemental Materials Table S3 and reliability estimates separately by gender are presented in Supplemental Materials Table S4) (Table 2).

We had relatively low missingness on study variables. Missingness on focal variables (HAB, HR, depression, anxiety, aggression) ranged from 0 to 8.7% with a median of 4.1%. Out of the 566 cases in our data set, 473 (84%) had complete data. Twenty-four participants (4.2%) were only missing values at Time 4, and twenty-one participants (3.7%) only had values at Time 1. No other pattern of missingness represented more than 2% of the cases. Our analyses do not provide us with reason to suspect the presence of non-random missingness in our data.

Figure 1A depicts the overall analytic model used to test the ability of HR to mediate the relations of HAB to aggression, anxiety, and depression symptoms. All analyses controlled for baseline levels of the outcome, child age at baseline, and child gender. The model for aggression had tolerable fit to the data, χ²(64) = 262.35, p <.001, CFI = 0.92, RMSEA = 0.07, 90% CI = (0.06, 0.08), SRMR = 0.04. HR significantly mediated the relation of HAB to parent-reported aggression. Higher levels of HAB at T1 were associated with higher levels of HR at T2, which were in turn associated with more parent-reported aggression at T3 (see Fig. 1B).

The model for depression had acceptable fit to the data, χ²(64) = 148.82, p <.001, CFI = 0.93, RMSEA = 0.05, 90% CI = (0.04, 0.06), SRMR = 0.04. HR did not significantly mediate the relation of HAB to parent-reported depression symptoms. Although HAB at T1 predicted HR at T2, HR did not predict depressive symptoms at T3 (see Fig. 1C).

The model for anxiety had tolerable fit to the data, χ²(64) = 202.77, p <.001, CFI = 0.92, RMSEA = 0.06, 90% CI = (0.05, 0.07), SRMR = 0.04. HR significantly mediated the relation of HAB to parent-reported anxiety symptoms. Higher levels of HAB at T1 were associated with higher levels of HR at T2, which were associated with more anxiety symptoms at T3 (see Fig. 1D).

Figure 2A depicts the overall analytic model used to test the ability of HAB to mediate the relations of HR to aggression, anxiety, and depression symptoms. The model for aggression had tolerable fit to the data, χ²(64) = 283.09, p <.001, CFI = 0.90, RMSEA = 0.07, 90% CI = (0.07, 0.08), SRMR = 0.04. HAB did not significantly mediate the relation of HR to parent-reported aggression. Unstandardized coefficients for individual pathways in this model are depicted in Fig. 2B. The model for depression had an acceptable fit to the data, χ²(64) = 181.64, p <.001, CFI = 0.90, RMSEA = 0.05, 90% CI = (0.05, 0.06), SRMR = 0.04. HAB did not significantly mediate the relation of HR to parent-reported depression symptoms (see Fig. 2C). The model for anxiety had acceptable fit to the data, χ²(64) = 177.71, p <.001, CFI = 0.93, RMSEA = 0.05, 90% CI = (0.04, 0.06), SRMR = 0.04. HAB significantly mediated the relation of HR to parent-reported anxiety symptoms; however, the direction of this indirect effect was negative (see Fig. 2D⁴).

It can be useful to assess parcel allocation variability, or the extent to which the analytic results vary depending on how the parceling was performed (Sterba & Right, 2016). We therefore reran our HAB through HR to psychological maladjustment models and our HR through HAB to psychological maladjustment models 100 times each, where we randomly assigned the HAB, HR, and psychological maladjustment items to each parcel. The variability in coefficients across these models is presented in Supplemental Material Table S5, whereas the percent of the random parcel models identifying significant mediation effects is presented in Table 3. Consistent with our primary analyses, a clear majority of the HAB through HR to aggression and the HAB through HR to anxiety random parcel models showed significant mediation, whereas a clear majority of the HAB through HR to depression symptoms models did not. Although we identified a significant HR through HAB to anxiety symptoms effect in our primary analysis, none of the HR through HAB to anxiety symptoms random parcel models showed significant mediation, suggesting that we should be cautious about drawing conclusions from this result. Consistent with our primary analyses, none of the HR through HAB to aggression random parcel models or the HR through HAB to depression symptoms random parcel models showed significant mediation.

We used multigroup analyses to determine whether the focal coefficients in our models varied significantly across our six samples separately for each outcome. These analyses were performed by comparing the fit of a model constraining the mediation coefficients and the direct effect to be equal across groups to a model allowing these coefficients to vary across groups. We did not find any evidence that these coefficients significantly varied across samples for the HAB through HR to outcomes models (all ps > 0.58) or for the HR through HAB to outcomes models (all ps > 0.22). An examination of changes in alternate fit indices also did not show evidence of substantial group differences (HAB through HR to outcomes: all ΔRMSEA ≤ 0.000, all ΔCFI ≤ 0.000, all ΔSRMR ≤ 0.001; HR through HAB to outcomes: all ΔRMSEA ≤ 0.000, all ΔCFI ≤ 0.003, all ΔSRMR ≤ 0.002; see Chen, 2007). Details of these analyses are provided in Supplemental Table S6.

We also used multigroup analyses to determine whether the mediation coefficients in our models significantly varied across male and female participants. Using the same procedures as for comparing samples, we did not find any evidence that the mediation coefficients and direct effects significantly varied across gender (HAB through HR to outcomes: all ps > 0.09, all ΔRMSEA ≤ 0.001, all ΔCFI ≤ 0.003, all ΔSRMR ≤ 0.003; HR through HAB to outcomes: all ps > 0.30, all ΔRMSEA ≤ 0.000, all ΔCFI ≤ 0.000, all ΔSRMR ≤ 0.002). Details of these analyses are provided in Supplemental Table S7.

An important limitation of our study was that HR was only assessed at a single point in time. When performing a longitudinal mediation analysis, best practices suggest that the model equations control for earlier assessments of the mediator and the outcome (Cole & Maxwell, 2003). Although our mediation models control for baseline assessments of the outcomes, we lacked the ability to control for baseline values of HR when it was acting as a mediator. Future investigators could provide more compelling causal evidence for the contributions of HR and HAB to psychological maladjustment if they examine the presence of mediation in a cross-lagged panel mediation model involving HR, HAB, and the outcomes, where both mediation models could be examined using the most appropriate controls.

Our study used self-report measures of HAB and HR that only capture youths’ explicit awareness of these cognitive processes. There are aspects of HAB and HR that occur beyond one’s explicit awareness and are not captured by self-report assessments (Prinstein et al., 2005). It is important for future studies to supplement self-report measures with assessments capturing implicit automatic processes such as selective attention to hostile information (Pollak & Tolley-Schell, 2003) or accessibility of hostile thoughts (Lindsay & Anderson, 2000). Another limitation to note is that although our measures of anxiety and depressive symptoms were theoretically derived from the CBCL, they were not validated CBCL scale scores. Finally, the reliability of HAB in our study was low, so we must be cautious about interpreting non-significant relations. Future research should focus on improving the assessment of HAB to enable clearer and more accurate interpretations.

Future studies of HAB employing community samples of youth might examine general beliefs not tethered to negative perceptions of others. One such example can be found in Yeager et al. (2013), who demonstrated that building beliefs that personality traits are flexible and changeable can reduce the likelihood youth ascribe negative intentions to others’ ambiguous behaviors. Reducing psychological maladjustment by targeting cognitive vulnerabilities such as HAB and HR might also be more effective for community samples of youth (e.g., Vassilopoulos et al., 2014) than youth with clinical levels of maladjustment (Hiemstra et al., 2019), indicating the importance of prevention. Corroborating our findings among youth with clinically elevated aggression, anxiety, and depression symptoms would be particularly informative for developing evidence-based treatments.

Additional contextual, cognitive, and emotional factors exist downstream from HAB and HR that we did not examine. For example, negative early life experiences with parents (e.g., chronic physical abuse) and peers (e.g., chronic social rejection) predict hostile and defensive patterns of processing social information as well as psychological maladjustment (Dodge et al., 2022). Other cognitive and emotional factors likely predict HAB and HR, such as irritability (Leigh et al., 2020). Future studies could examine whether irritability mediates the relation of HR to psychological maladjustment and assess the utility of a serial mediation model suggesting that HAB influences HR, which leads to irritability, which finally leads to psychological maladjustment. Existing evidence supports such a model. For example, irritability predicts internalizing and externalizing symptoms in adolescence (Sorcher et al., 2022), is associated with biases toward anger (Salum et al., 2017), and is associated with HR (Caprara et al., 2007).