Descriptive statistics
There was a significant difference in the waiting time between the pre-treatment assessment and the first treatment session, t (56.77) = -4.15. Participants in the ACT condition (M = 3.70, SD = 2.60) had a shorter waiting time (in weeks) compared to participants in the CBT intervention (M = 7.18, SD = 4.57, p <.001). We therefore entered this variable as a covariate in the main analyses. There were no significant differences between conditions at the pre-treatment assessment (ps = .12 to .97) on any of the other measures. More information can be found in A-Tjak et al. [7].
Long term outcome
Depressive symptom scores (QIDS-SR) and scores of quality of life (EUROHIS) are presented in Table 1. Results from post-treatment and 6-month follow-up have already been reported in our previous article [7]. We present them here for the convenience of the reader. The results of multilevel analyses can be found in Table 2. Patients reported large reductions in depressive symptoms from pre-treatment to 12-month follow-up. There were no differences in the rate of change between ACT and CBT from 6-month follow-up to 12-month follow-up, as indicated by a non-significant Cubic Time × Condition interaction. Patients also reported large increases in quality of life from pre-treatment to 12-month follow-up. The Cubic Time × Condition interaction was not significant, indicating no differences between ACT and CBT in the rate of change in quality of life from 6-month follow-up to 12 month follow-up.
Mediation
Depressive symptoms, dysfunctional attitudes, decentering and experiential avoidance scores at pre-treatment, during treatment and post-treatment are presented in Table 3. Overall model fit information for the models that were fitted are presented in Table 4.
Depressive symptoms. Model fitting procedures for the univariate LGCMs indicated a poor fit for the intercept-only model and a good fit for the linear model. A comparison of the BIC values suggests that the linear change model provided the best fit. This indicates that the assumption of linear change in depressive symptoms over time is preferred over the assumption of no change and curvilinear change. Parameter estimates of the linear change model indicated a significant linear decrease, M = -1.58, p < .001, in depressive symptoms at the group level. The variances of the intercept, 15.79, and the slope, 1.02, were significant, p < .01, showing significant individual differences of the initial depressive symptom scores and the rate of change over time. Using a multigroup approach on the linear model, we tested whether the trajectory in the group receiving ACT differed from the trajectory in the group receiving CBT. Inspection of the BIC values indicated that the fit of the model in which the slope factor means were allowed to be estimated differently in the two groups was inferior to the fit of the model in which the slope factor means were constrained to be equal. This implies that the depressive symptom linear slopes were not significantly different from one another.
Dysfunctional attitudes. Model fitting procedures for the univariate LGCMs indicated a poor fit for the intercept-only model and an acceptable to good fit for the linear and curvilinear model. A comparison of the BIC values indicates that the quadratic change model provided the best fit. This indicates that the assumption of curvilinear change in dysfunctional attitude scores over time is preferred over the assumption of no change and linear change. Parameter estimates of the curvilinear change model indicated a significant linear decline, M = -1.35, p < .001, in dysfunctional attitudes at the group level. The mean of the quadratic slope was -.26, p = .045, suggesting an acceleration in the rate of decrease over time. Variances of the intercept, 286.32, and linear slope, 1.22, were both significant, p < .001, showing significant between-person variance of the initial dysfunctional attitudes scores and the rate of change over time. The variance of the curvilinear slope was nonsignificant, .25, p = .34, suggesting no between-person variability in the reversal of the rate of change. Using a multigroup approach on the curvilinear model, we tested whether the trajectory in the group receiving ACT differed from the trajectory in the group receiving CBT. Inspection of the BIC values indicated that the fit of the model in which the slope factor means were allowed to be estimated differently in the two groups was inferior to the fit of the model in which the slope factor means were constrained to be equal. This implies that the dysfunctional attitudes linear and quadratic slope were not significantly different from one another.
In step 2, the relations among change in dysfunctional attitudes and depressive symptoms were assessed in a parallel process model. Providing support for common developmental trends in dysfunctional attitudes and depressive symptoms, the dysfunctional attitudes slope was significantly associated with the slope in depressive symptoms, as indexed by a regression coefficient of .28, p < .01. In addition, the intercepts of dysfunctional attitudes and depressive symptoms were significantly associated, .09, p = < .01.
In step 3, a latent difference score model was fitted. See Figure 1 for parameter estimates of this model. For each interval, depressive symptom levels at an earlier time point significantly predicted depressive symptom levels at a later time point. Above and beyond these autoregressive effects, each of the dysfunctional attitudes difference score factors significantly predicted symptom levels of depression at the end of the interval. This result indicates that prior change in dysfunctional attitudes was related to subsequent symptom levels of depression for each of the intervals. A multigroup model in which the difference score factors predicting levels of depression were allowed to be estimated differently for ACT and CBT resulted in an inferior fit, relative to a model in which the regression pathways were constrained to be equal across treatment conditions. Although the fit of both multigroup models was not optimal, this result indicates that, in contrast to what we predicted, the assumption of a non-specific treatment effect of change in dysfunctional attitudes on depressive symptom levels is preferred over the assumption of treatment-specific effects.
[Figure 1 here.] Figure 1: Latent difference score model with dysfunctional attitudes as the mediator and depressive symptoms as outcome. Path coefficients (unstandardized) are aggregated across conditions (ACT, CBT) as this model resulted in the best fit. DAS = DAS-17 scores, QIDS = QIDS-SR scores, pre = pre-treatment, s = session number during treatment, post = post-treatment, ∆ = latent difference score.
*** p < .001, **p < .01, *p <.05.
Decentering. Model fitting procedures for the univariate LGCMs indicated a poor fit for the intercept-only model and the linear change model. The quadratic change model had a considerably better fit, indicating that the assumption of curvilinear change in decentering over time is preferred over the assumption of no change and linear change. Parameter estimates of the curvilinear change model indicated a significant linear increase, M = .30, p < .01, in decentering at the group level. The mean of the quadratic slope was .25, p < .001, suggesting an acceleration in the rate of increase over time. Variances of the intercept, 24.91, linear slope, .38, and quadratic slope, .14, were significant, p < .05, showing significant between-person variance of the initial decentering scores, the rate of change over time and the reversal in the rate of change. Using a multigroup approach on the curvilinear model, we tested whether the trajectory in the group receiving ACT differed from the trajectory in the group receiving CBT. Inspection of the BIC values indicated that the fit of the model in which the slope factor means were allowed to be estimated differently in the two groups was inferior to the fit of the model in which the slope factor means were constrained to be equal. This implies that the decentering linear and quadratic slope were not significantly different from one another.
In step 2, the relations among change in decentering and depressive symptoms were assessed in a parallel process model. Providing support for common developmental trends in decentering and depressive symptoms, the decentering slope was significantly associated with the slope in depressive symptoms, as indexed by a regression coefficient of -.71, p < .01. In addition, the intercepts of decentering and depressive symptoms were significantly associated, as indexed by a regression coefficient of -.37, p = < .01.
In step 3, a latent difference score model was fitted. See Figure 2 for parameter estimates of this model. Above and beyond the significant autoregressive effects, the decentering difference score factors for the first and third interval significantly predicted symptom levels of depression at the end of these intervals. Across treatment groups, this result indicates that changes in decentering from pre-treatment to session 6, and from session 6 to 11 were related to subsequent symptom levels of depression, whereas no evidence for such an effect was found for the remaining intervals. However, a multigroup model in which the difference score factors predicting levels of depression were allowed to be estimated differently for ACT and CBT resulted in a superior fit, relative to a model in which the regression pathways were constrained to be equal across treatment groups. Although the fit of both multigroup models was not optimal, inspection of the pathways for ACT shows that changes in decentering from pre-treatment to session 1, and from session 6 to 11 significantly predicted subsequent symptom levels of depressive symptoms. For CBT, changes in decentering from pre-treatment to session 1, and from session 1 to 6 significantly predicted subsequent symptom levels of depression. In contrast to what we expected, these results indicate a non-specific treatment effect of change in decentering on symptom levels of depression.
[Figure 2 here.] Figure 2: Latent difference score model with decentering as the mediator and depressive symptoms as outcome. Path coefficients (unstandardized) are estimated separately for ACT (italics) and CBT (boldfaced) as this model resulted in a superior fit to a model with equality constraints on the difference score to outcome path coefficients. EQ-D = EQ-D scores, QIDS = QIDS-SR scores, pre = pre-treatment, s = session number during treatment, post = post-treatment, ∆ = latent difference score.
*** p < .001, **p < .01, *p <.05, ns = non-significant.
Experiential avoidance. Model fitting procedures for the univariate LGCMs demonstrated a poor fit for the intercept-only model and a good fit for the linear and curvilinear model. Negative residual variances were obtained in the curvilinear model, suggesting that the assumption of curvilinear growth may not be reasonable for the observed data. This indicates that the assumption of linear change in experiential avoidance scores over time is preferred over the assumption of no change and quadratic change. Parameter estimates of the linear change model indicated a significant linear decrease, M = 1.38, p < .001, in experiential avoidance at the group level. The variance of the intercept was significant, 23.67, p < .001, whereas the variance of the slope was nonsignificant, 1.38, p = .06, indicating significant individual differences of the initial experiential avoidance scores but no individual differences in the rate of change. Using a multigroup approach on the linear model, we tested whether the trajectory in the group receiving ACT differed from the trajectory in the group receiving CBT. Inspection of the BIC values indicated that the fit of the model in which the slope factor means were allowed to be estimated differently in the two groups was inferior to the fit of the model in which the slope factor means were constrained to be equal. This implies that the experiential avoidance linear slopes were not significantly different from one another.
In step 2, the relations among change in experiential avoidance and depressive symptoms were assessed in a parallel process model. Providing support for common developmental trends in experiential avoidance and depressive symptoms, the experiential avoidance slope was significantly associated with the slope in depressive symptoms, as indexed by a regression coefficient of -.92, p < .05. In addition, the intercepts of experiential avoidance and depressive symptoms were significantly associated, as indexed by a regression coefficient of -.31, p = < .001.
In step 3, a latent difference score model was fitted. See Figure 3 for parameter estimates of this model. Above and beyond the significant autoregressive effects, the experiential avoidance difference score factors for the last interval significantly predicted symptom levels of depression at the end of these intervals. Across treatment groups, this result indicates that changes in experiential avoidance from session 16 to post-treatment were related to subsequent symptom levels of depression, whereas no evidence for such an effect was found for the remaining intervals. However, a multigroup model in which the difference score factors predicting levels of depression were allowed to be estimated differently for ACT and CBT resulted in a superior fit, relative to a model in which the regression pathways were constrained to be equal across treatment groups. Although the fit of both multigroup models was not optimal, inspection of the pathways for ACT shows that changes in experiential avoidance for all of the intervals, except the first and second interval (pre-treatment to session 1, and session 1 to 6), significantly predicted subsequent symptom levels of depressive symptoms. For CBT, changes in experiential avoidance for all of the intervals did not significantly predict subsequent symptom levels of depression. This result is in line with the expectations, and indicates a specific treatment effect of change in experiential avoidance on depressive symptoms levels for ACT.
[Figure 3 here.] Figure 3: Latent difference score model with experiential avoidance as the mediator and depressive symptoms as outcome. Path coefficients (unstandardized) are estimated separately for ACT (italics) and CBT (boldfaced) as this model resulted in a superior fit to a model with equality constraints on the difference score to outcome path coefficients. AAQ = AAQ-II scores, QIDS = QIDS-SR scores, pre = pre-treatment, s = session number during treatment, post = post-treatment, ∆ = latent difference score.
*** p < .001, **p < .01, *p <.05, ns = non-significant.
[Table 4 here]