CAPÍTULO II: EXPERIMENTO Y MÉTODOS 83
II.3 Procesado de los datos 96
II.3.2 Cámara hemisférica 98
We used Mplus 7.3 (Muthén & Muthén, 1998–2015) for all analyses in Study II. We used R 3.4.3 (R Core Team, 2017) for all analyses in Studies I and IV and some analyses in Study III. We further used SPSS 24 for some analyses in Study III. In Study IV, we used the mice 2.9 R package (van Buuren & Groothuis-Oudshroon,
2011) for multiple imputation and the lavaan 0.6-3 (Rosseel, 2012) and semTools 0.5-1 R packages (Jorgensen, Pornprasertmanit, Schoemann, & Rosseel, 2018) to allow for maximum-likelihood path analyses in multiply imputed datasets. For linear mixed- effects modeling in Study III, we used the nlme 3.1-137 R package (Pinheiro, Bates, DebRoy, Sarkar, & R Core Team, 2018). To increase transparency, we retained all R input scripts used and will make them freely available to any interested parties.
7.3.1
Study I
We calculated descriptive statistics on the characteristics and quality of the included studies. We assessed the relationship between year of publication and study quality by a correlation analysis.
7.3.2
Study II
We used linear multiple regression to examine the effects of TRT on PTCs at posttreatment and follow-up, controlling for pretreatment levels. We then used the Wald test to examine gender effects by testing equality constraints for regression coefficients for boys and girls.
For indirect effects, we used maximum-likelihood path analysis based on the correlation matrix to examine mediated effects of the intervention on PTSS at follow-up via effects on PTCs at posttreatment. We used the product of coefficients method with a second-order Taylor series approximation for the standard error to quantify and examine the mediated effect (MacKinnon et al., 2002).
We used latent class growth analysis (Berlin, Parra, & Williams, 2014) to detect trajectories in the level and change of the children’s PTCs during the intervention. We estimated linear latent class growth analysis models and selected the best-fitting model based on the Bayesian information criterion and the bootstrap likelihood ratio test (Nylund, Asparouhov, & Muthén, 2007). We then used multinomial logistic regression to study whether gender, age, depressive symptoms, PTSS, or number of traumatic war experiences would predict most likely membership in any particular trajectory (Clark & Muthén, 2009).
We tested two different latent measurement models for PTSS and two for PTCs. However, as clear measurement invariance could not be established between the intervention and control groups, we finally used observed total scores on the CRIES and CPTCI measures instead.
We used a form of Kish correction1 (Ukoumunne, Gulliford, Chinn, Sterne, &
Burney, 1999) to correct unrealistically small estimates of population standard errors that might have resulted from the cluster sampling method and associated non- independence of observations. We used full-information maximum likelihood procedures to account for missing data.
7.3.3
Study III
In primary analyses, we used repeated-measures analyses of variance to examine changes in total PTSS, intrusion symptoms, avoidance symptoms, and arousal symptoms from pretreatment to posttreatment in treatment completers. We set time as a two-level within-subject variable and treatment as a two-level between-subjects variable and examined main and interaction effects. We used similar analyses for depressive symptoms, resilience, and self-evaluated and guardian-evaluated psychological distress. Post hoc power analyses indicate that these analyses were only powered to detect large effects of around f = 0.5 at 80% power. We then used t tests to examine the significance of within-group changes in total PTSS, intrusion symptoms, avoidance symptoms, and arousal symptoms, for the NET and TAU groups separately.
For assessing clinical significance of changes in PTSS, we used the cut-off score of 17 for the avoidance and intrusions subscales of the CRIES to examine how many participants in each group exceeded this cutoff and thus had probable PTSD at pretreatment and posttreatment. We then used related-samples McNemar tests to check whether within-group changes in the share of probable PTSD were significant. In intention-to-treat analyses, we used linear mixed-effects modeling employing all available data from all participants at all points of measurement. We modelled the effect of time as weeks elapsed since the pretreatment assessment until each point of assessment. We employed a stepwise model building approach, where we allowed for random intercepts, random slopes, and autocorrelation one by one and retained them if they improved model fit at each step. At the last step, we checked whether adding Time × Treatment interactions would improve model fit and be statistically significant. We assessed improvements in model fit by the Akaike information
1 In this correction, standard errors of regression parameters in all regression-based analyses were multiplied by the square root of the design effect, 1 + (n − 1) * ICCx * ICCy, where ICCx is the intraclass correlation coefficient of the covariate and the ICCy the intraclass correlation coefficient of the dependent variable, and n is the average size of the clusters, in this case the average number of
criterion and likelihood ratio tests. We computed 95% confidence intervals for all effects. We used visual inspection of residual plots to check for evidence of heteroscedasticity or significant deviations from normality, but found none.
Individual missing item-level responses in otherwise completed measures (a total of 34 data points) were replaced by the response closest to that participant’s mean answer to other items of the measure. For entirely missing measures, we used pairwise deletion for each analysis, except for intention-to-treat analyses where we used all available data.
7.3.4
Study IV
We used repeated-measures analyses of variance to assess changes in negative PTCs and quality of traumatic memories over the course of treatment and possible differences between NET and TAU in such changes. Time was set as a three-level within-subject factor (pretreatment, midtreatment, posttreatment) and treatment as a two-level between-subjects factor, and we examined main and interaction effects.
We assessed the relationship between pretreatment to posttreatment changes in negative PTCs and quality of traumatic memories and PTSS by correlation analyses. Finally, for mediation analysis, we used maximum-likelihood path analysis, with separate path models for negative PTCs and quality of traumatic memories. In each model, PTSS at posttreatment were regressed on PTSS at pretreatment, a dummy variable for type of treatment, and PTC/traumatic memory scores at posttreatment. The PTC/traumatic memory scores at posttreatment were then regressed on their values at pretreatment as well as PTSS at pretreatment and the dummy variable for type of treatment. We used the product of coefficients method to estimate indirect effects of treatment on PTSS at posttreatment via PTCs and traumatic memories at posttreatment, and the conservative asymptotic normal distribution method to construct confidence intervals around the estimate to assess the significance of the indirect effect (MacKinnon et al., 2002).
We used multiple imputation by chained equations to account for the notable amount of missing data. We generated fifty imputed data sets to replace missing data at the item level (Gottschall, West, & Enders, 2012). We considered all available variables used in the study as potential predictors, as well as available demographic variables, data on depressive symptoms, psychological distress, and assessments of PTSS, PTCs, and quality of traumatic memories at follow-up, where available. We included as predictor variables for multiple imputation those variables that correlated
at .30 or more with the variable being imputed. We carried out all analyses with these multiple imputed data sets, accordingly marked with an MI subscript. Due to large share of missing data at midtreatment and associated imprecision in imputing them, we only used midtreatment data as additional data points in repeated measures analyses of variance and not in mediation analyses.
7.4
Ethical considerations
The board of directors of the Gaza Community Mental Health program and officials in the government of Gaza approved the research for Study II. The ethical boards of Pirkanmaa Hospital District, Tampere City Welfare Services, Helsinki Diaconess Institute, and the Hospital District of Southwest Finland approved Studies III and IV.
Age-appropriate information leaflets about the study were prepared and given to potential participants and their guardians in Studies III and IV. Information sheets about the study and its purpose were also provided to children’s parents in Study II. Participation in all studies was voluntary. Oral consent to participate was received from participants and their parents in Study II. Written consent to participate was received from both the participants themselves and their parents or guardians in Studies III and IV. In the NET trial, we performed interim analyses roughly half way through the recruitment period to ensure that NET was not performing worse than TAU to a clinically significant degree.
The authors had no conflicting interests at the time of the studies. Since the completion of Studies III and IV, the authors of these studies have received modest income from organizing clinical training in NET.