Childhood Trauma Exposure
Trained psychologists interviewed children with the life-events checklist (LEC) during T2. This semi-structured checklist is part of the Clinician Administered PTSD Scale for Children and Adolescents (CAPS-CA) (van Meijel et al. 2019
; Pynoos et al., 2015
). The checklist consists of 25 possible traumatic items, with five answer options; ‘Happened to me’
, ‘Witnessed it
’, ‘Learned about it
’, ‘Not sure
’ and ‘Doesn’t apply
’. The LEC has good psychometric properties with a test–retest reliability of r
= 0.82 and convergence validity with a mean kappa for all items of 0.61 (Gray et al., 2004
). Besides questioning exposure to events, we also asked the child’s age during these events. After data collection and data entry, we rated the traumatic events based on the DSM–IV criteria. All events were rated by at least 2 out of 3 independent coders (RodK; JE; HB) and discrepancies were resolved by discussion with a third coder or expert panel.
We decided to include events that concerned participants themselves, their first degree relatives, or best friends only. Exceptions for this decision were cases of extreme violence such as victims of the attack on flight MH17 (in that case also teachers and friends were included). As a traumatic event is defined as one involving an actual or threatened death, serious injury or sexual violation, we decided that only severe accidents where an ambulance or hospital stay was needed, were included. For domestic emotional abuse we only included events that were extreme or caused structural safety issues for a longer period of time. As emotional abuse is mostly vaguely described by children (e.g. by mother yelled at me/called me names), we only included this as a traumatic event when the children reported that this happened more than once over a longer period of time. This was done by a “blind” expert panel of five experts in the field. In cases of discrepancies, consensus was reached by discussion. Approximately one-third of the children had been exposed to a traumatic event. Most traumatic events were severe accidents (27.5%) and victim of community violence (22.5%). Other events were disaster (1.6%), victim of domestic violence (7.8%), witness of domestic violence (7.5%), witness of community violence (5.9%), sexual assault (2.2%), death or injury of a loved one (6.7%), serious medical condition (14.9%), and other events (3.5%). Based on the interviews, we constructed two variables: traumatic events until the age of 5 (no / ≥ 1 event), and traumatic events between 6 and 12 years old (no, 1 event, ≥ 2 events). Twenty children (2.0%) did not participate in the interview.
Based on earlier research and theories, our approach was to focus on the conceptual unity underlying different aspects of executive functioning (Miyake & Friedman, 2000
; Diamond, 2013
). At T1, executive functioning was measured using subtests of the widely-used Amsterdam Neuropsychological Tasks (ANT) (Sonneville, 1999
). The ANT is a computerized test battery that was performed in an individual setting at school in which the children performed the tasks pursuit, tracking, and response organization objects (ROO). The tasks have been shown to be sensitive to detection of neuropsychological problems in various samples and have good reliability and validity (De Sonneville et al., 2002
; Rowbotham et al., 2009
).The ROO task measures inhibitory control and cognitive flexibility and consists of three parts that increase in complexity. In part 1, children had to click the left mouse button when a green ball appeared on the left side of the screen and vice versa. In part 2, the tasks requires a click on the right mouse button when a red ball appeared on the left side of the screen and vice versa. In part 3, children had to follow these instructions based on the color of the ball that randomly alternated. A valid response was considered when a child clicked the correct button between 200 to 6000 ms after the stimulus was presented on the screen. Both the pursuit and tracking task measure visuomotor coordination. In the pursuit task, the child had to follow a mouse cursor on the screen that made a random trajectory with a constant speed of 10 mm/s, using their non-preferred hand and in the second part with their preferred hand. The tracking task is similar to the pursuit task, but in this task the mouse cursor follows a familiar and planned trajectory, which requires less executive demands.
The following outcome measures of these tasks were used to assess executive functions at age 5 (1) flexibility 1: mean reaction time compatible part 3 minus mean reaction time compatible part 1 in milliseconds, (2) flexibility 2: number of errors compatible part 3 minus mean reaction time compatible part 1, (3) flexibility 3: standard deviation right plus left hand compatible part 3 minus standard deviation right plus left hand compatible part 1 in milliseconds, (4) inhibition 1: mean reaction time incompatible part 2 minus mean reaction time compatible part 1 in milliseconds, (5) inhibition 2: number of errors incompatible part 2 minus number of errors number of errors compatible part 1, (6) inhibition 3: standard deviation right plus left hand compatible part 2 minus standard deviation right plus left hand compatible part 1 in milliseconds, and (7) motor flexibility: mean deviation overall pursuit – mean deviation overall tracking. The variables included were those that were most often reported focusing on inhibition and flexibility (Guxens et al., 2016
; Menting et al., 2018
). There was some missing data for these variables, as 15.51% of the children did not participate in the tasks. Furthermore, as the outcome is assumed to be unreliable when children outperform the difficult trials compared to the control trials of the tasks, negative contrast scores (in 0% to 14.5% of the cases) were recoded as invalid. To improve model convergence, we divided the values by constants to obtain variances with values between 1 and 10. A higher score on a variable corresponds with worse executive functioning.
We examined whether these variables could be modeled to load on one latent factor for executive functioning using the maximum likelihood with robust standard errors (MLR) estimator. Step 1 was to load the seven variables on one latent variable. This model had a poor model fit (Χ2
(14) = 462.69, p
= 0.00, CFI = 0.69. RMSEA = 0.19). Modification indices showed – step by step – that the errors of inhibition 3 and flexibility 1, inhibition 1 and 3 should covary to improve the model. However, after adding the last error covariance, this model did not converge due to negative residual variance of inhibition 1. As this residual variance was non-significant, we could constrain the residual variance to zero and did not add the error covariance. We continued with adding step by step error covariances based on the modification indices between inhibition 3 and flexibility 3, flexibility 1 and 2, inhibition 2 with flexibility 2, inhibition 2 with flexibility 1, and inhibition 2 and 3. Inspection of the factor loadings indicated that inhibition 2 (0.11, p
= 0.39) and flexibility 2 (0.17, p
= 0.17) both had non-significant factor loadings on the latent variable. We excluded these variables (and their added error covariances) from the model. Therefore, the final measurement model included flexibility 1, flexibility 3, inhibition 1, inhibition 3, and motor flexibility as shown in Fig. 1
. This model had excellent fit (Χ2
(4) = 0.38, p
= 0.99, CFI = 1.00. RMSEA = 0.00), with standardized factor loadings ranging between 0.21 for motor flexibility to 0.76 for flexibility 3.
To measure executive functions at T2, 24 items (of a total of 75 items) of the Dutch parent version of the Behavior Rating Inventory for Executive Functioning (BRIEF) were used (Gioia et al., 2001
; Huizinga & Smidts, 2009
). The selected items cover eight subscales with three items each which were rated by caregivers. The questionnaire has eight subscales (inhibit, shift, emotional control, initiate, working memory, plan/organize, organization of materials and monitor) that are covered by the two indices Behavior Regulation Index (BRI) and Metacognition Index (MI). Statements such as “he/she struggles with finishing tasks” and “he/she gets upset in new situations” are scored on a three-point scale (1
= never, 2
= sometimes, 3
). This means that a higher score on the subscales indicate poorer executive functioning. The questionnaire showed good psychometric properties in a sample of parents of 847 children with Cronbach’s alpha’s ranging from 0.78 to 0.96 (Huizinga & Smidts, 2009
). Due to the long battery of questionnaires and to decrease the burden of participating in the research, we selected 24 items. In our study, the 24 items version of the BRIEF had an excellent reliability on item-level, as indicated by a Cronbach’s alpha of 0.91. Of the 1006 participants, 55 participants (5.5%) did not fill out the questionnaire. It is important to note that a higher score on these items corresponds with worse executive functioning.
We also examined whether these subscales could be modeled to load on one latent variable for executive functioning at age 12 using the MLR estimator. A model with all eight subscales of the BRIEF loading on one latent factor with error variances allowed to covary based on step-by-step modification indices, had an excellent model fit (Χ2 (7) = 8.79, p = 0.27, CFI = 0.99, RMSEA = 0.016). Error variances that covaried were: planning with initiate; emotion regulation with flexibility; inhibition with behavior evaluation, emotion regulation, and flexibility; behavior regulation with emotion regulation and flexibility; initiate with flexibility; organizing with working memory, flexibility and initiate. All standardized factor loadings were significant and in the expected direction and ranged from 0.31 for flexibility to 0.89 for working memory.
Parents reported on their parenting behavior by filling out the shortened version (32 items) of the Parental Styles and Dimensions Questionnaire (PSDQ) at T1. This scale was developed to investigate parenting styles using specific parenting practices that occur within the authoritative, authoritarian, and permissive parenting style. Due to the long battery of questionnaires and to decrease the burden of participating in the research, we used the shortened version of the questionnaire. This version consists of 15 items in the authoritative scale, 12 in the authoritarian scale, and 5 in the permissive parenting scale. In our study, items were rated on a four-point Likert type scale (1
= (almost) never; 2
= once in a while; 3
= often; 4
for readability of the overall test battery. Parents responded on questions such as “I encourage my child to talk about its troubles”, “I punish by taking privileges away from my child with little if any explanation”, and “I spoil my child”. Scales were calculated by taking the sum score of the items within that scale. Psychometric properties of the 32-PSDQ have been investigated across various studies. Cronbach’s alphas ranged between 0.82 and 0.91 for authoritative, 0.67 and 0.86 for authoritarian, and 0.58 and 0.79 for permissive parenting. Good concurrent and predictive validity was also reported (Olivari et al., 2013
). Although validity research on the shortened version is scarce, one study found its concurrent validity in relation to three other questionnaires to be sufficient (Topham et al., 2011
). In our sample, we found Cronbach alpha’s of 0.82 for authoritative parenting, 0.71 for authoritarian parenting, and 0.59 for permissive parenting. As we found the reliability of the permissive parenting scale to be insufficient, we did not include these in our analyses. To assess the moderating role of parenting behavior, we used multi-group analyses. Therefore, the sample was split across the median for each of the parenting dimensions to create equal groups (authoritative parenting: 16; authoritarian parenting: 8).
To answer our research questions, we performed structural equation modeling (SEM) using Mplus 7 (Muthén & Muthén, 2012
) for analyses. Little’s Missing Completely at Random (MCAR) test was significant (Χ2
(292) = 541.30, p
= 0.000), therefore we assumed that data were not missing completely at random. As missingness was not predictable from the dependent variables, we assumed that the data was Missing At Random (MAR) (Tabachnick & Fidell, 2013
). We investigated whether cases with or without any missing data were significantly different from each other on all included variables using independent T-tests. Independent T-tests did not show significant differences between participants with missing data on T1 on measures of executive functioning at T2 nor the other way around. However, we found significant differences for all outcome measures of the ROO task and for authoritarian parenting behavior. This means, that on these variables, the mean scores were different for participants that had no missing data compared to participants that had missing information on one of the variables of interest. For prenatal substance abuse, we found significant differences between our sample and the total sample at birth (X2
(1) = 29.23, p
= 0.00) as more mothers reported on prenatal substance use in our sample. For trauma exposure and executive functioning at age 12, we were not able to check whether our sample differed from the total sample of the birth cohort (starting at birth) as we only included participants that reported on trauma exposure at age 12. We checked normality of the data by investigating skewness and kurtosis and divided these statistics by their standard error. For all executive functioning variables, we found extreme positive skewness and kurtosis, which improved after dealing with univariate outliers. We modified the values to the closest observed value plus or minus one unit when z-scores exceeded ± 3.29, which resulted in an improved — but non-normal — distribution. We did not transform variables, as this would make interpretation merely impossible. We ran all models also with censored variables, and differences in models are reported when this was the case. We used the weighted least squares means and variance adjusted (WLSMV) estimator for the model analyses.
We constructed a longitudinal model as depicted in Fig. 1
. After running our hypothesized model, we used multi-group analyses to investigate whether the link between trauma exposure and executive functioning was different across relatively low and high parenting behavior along the dimensions of authoritarian, authoritative, and permissive parenting. Model fit was assessed using comparative fit index (CFI; good model fit > 0.90) and Root Mean Square Error of Approximation (RMSEA; good model fit < 0.08) (Kline, 2005