Moderation of treatment effects by parent-adolescent conflict in a randomised controlled trial of Attachment-Based Family Therapy for adolescent depression

Because parent-adolescent conflict has shown a consistent association with the development and course of adolescent depression, it has also been studied as a potential predictor or moderator of outcome in multiple clinical trials of adolescent depression treatments. Moderators in the context of clinical trials have been defined as baseline variables across which the effect of treatment allocation on treatment outcome varies. Baseline variables that are associated with treatment outcome independently of treatment allocation are referred to as non-specific predictors (9).

Adolescent-reported parent-adolescent conflict was found to be a non-specific predictor of both treatment nonresponse and depression recurrence in a trial of different psychotherapies for adolescent depression (10). In the Treatment of Selective Serotonin Reuptake Inhibitor-Resistant Depression in Adolescents Study, adolescent report of more parent-adolescent conflict was also found to be a non-specific predictor of nonresponse (11). Mother-report of frequent and intense conflict was similarly found to be a non-specific predictor of poor outcomes in an exploratory analysis of data from the Treatment of Adolescent Depression Study (12).

However, in a trial comparing Interpersonal Therapy to treatment as usual, adolescents reporting more conflict with mothers at baseline benefited more from Interpersonal Therapy, which is a moderator effect (13). Similarly, in a preventive group intervention based on Interpersonal Therapy with school counselling, the intervention was superior to school counselling only for those adolescents who reported heightened parent-adolescent conflict (14), also showing moderation by parent-adolescent conflict for the interpersonally focused treatment. Another comparable finding was reported from a trial comparing family-focused treatment to enhanced usual care for adolescents with a bipolar disorder (15). For adolescents from families reporting heightened expressed emotion, the family focused treatment was superior to enhanced usual care, but this was not the case for adolescents from families with lower levels of expressed emotion. Seen together, these findings suggest that while parent-adolescent conflict can impede treatment, treatments gains can perhaps be made in these cases by focusing treatment on family-related issues, making treatment effectiveness conditional on the level of conflict or family distress.

Attachment-Based Family Therapy (ABFT) is a manualised family therapy for adolescent depressive symptoms and suicidal ideation (16). ABFT initially focuses on identification and repair of relational ruptures between depressed adolescents and their parents. Building on reduced conflict and renewed trust in the parent-adolescent relationship, the family is then guided in collaborative work to reduce depressive symptoms and improve functioning. The developers of the intervention have conducted several clinical trials and other program evaluations (17), and ABFT has been designated a probably efficacious treatment (18). Still, in the largest randomised controlled trial conducted to date, ABFT was not found to be superior to family-enhanced non-directive supportive therapy for reducing adolescent suicidal ideation (19). A secondary analysis of that trial found that observations of less cooperative family communication, as well as non-white race and lower income-to-needs ratio predicted higher treatment benefit in both trial arms (20).

In Norway, an initial study found ABFT implementation in Norwegian public child and adolescent mental health services to be feasible and the treatment to be acceptable to Norwegian families (21). A larger Norwegian randomised controlled trial comparing 16 weeks of ABFT to treatment as usual (TAU) for adolescent depression (clinicaltrials.gov identifier NCT01830088) was conducted to follow up on these findings. Contrary to the primary hypothesis of the trial, ABFT was not found to be superior to treatment as usual (22). While the findings from these recent trials do not provide evidence that ABFT on average is more effective than treatment as usual or other active comparisons in treating adolescent depression or suicidal ideation, ABFT is a treatment where moderation of effectiveness by parent-adolescent conflict is highly plausible. In line with this, we evaluate the moderator hypothesis that the difference in outcome between ABFT and treatment as usual would be larger at higher levels of parent-adolescent conflict and in favour of ABFT.

Participating families were recruited among adolescents referred to two Child and Adolescent Mental Health Services (CAMHS) in Southe-Eastern Norway. The clinics were publicly funded, and all treatments were provided free of charge to the patients and their families, within the framework of the universal health insurance system of Norway. During pre-specified recruitment periods, all referral letters for adolescents (13 - 17 years) were examined for mentions of depression or core depressive symptoms (depressed mood, anhedonia, or fatigue). The CAMHS routinely administered the Youth Self Report (23), and these were screened for raw scores on the Affective Problems subscale above 6 to find depressed adolescents not identified as such in their referral letters (24). Eligible adolescents or their parents, depending on adolescent age, were contacted by telephone and invited to participate in a randomised trial of family therapy for adolescent depression. Participants were required to be currently living with an adult who had become a caregiver for them before age four, and willing to have this adult participate in treatment. Interested adolescents meeting these criteria were screened with Beck Depression Inventory-II (25) over telephone and invited for an assessment session if they scored above 17, a threshold expected to maximise sensitivity (26). Adolescents were included in the study if they scored above 15 on the Grid Hamilton Depression Rating scale (GRID-HAMD, 27) and met diagnostic criteria for a current Major Depressive Disorder (28). Adolescents meeting criteria for a psychotic disorder, anorexia nervosa, bipolar disorder, intellectual disability or pervasive developmental disorder were excluded from the study. In a small number of cases, exclusionary criteria (psychotic disorder or atypical anorexia nervosa) were not detected at baseline but uncovered during treatment. One family withdrew consent shortly after randomisation and are not included in any analyses. Sixty participants were randomised, of which 52 (87%) were female. Figure 1 shows the flow of participants through the study.

FIGURE 1.

Participating adolescents and their parents met with a study-affiliated clinical psychologist (the first or second author) at the CAMHS and written informed parental consent and adolescent assent was obtained. Adolescents and parents were then interviewed separately with the Kiddie Schedule for Affective Disorders and Schizophrenia - Present and Lifetime version (K-SADS-PL, 29), and the depressive symptoms of the adolescent were further assessed with the GRID-HAMD (27). All interviews were video-recorded. Both parents and adolescents completed self-report measures of parent-adolescent conflict before randomisation. If the adolescent met inclusion criteria, the assessing clinician conducted randomisation by opening a sealed, numbered envelope containing the treatment allocation. Randomisation was stratified by site, age (13-15 years and 16-17 years), gender (male and female), and severity of depression (GRID-HAMD score of ≤ 24 and ≥ 25). Parents and adolescents were given feedback on diagnosis and treatment allocation at the end of the assessment session. The assessing clinician answered questions from parents and the adolescent concerning the assessment and implemented standard safety monitoring procedures to the extent deemed necessary. CAMHS staff were then informed of treatment allocation and given a report of the assessment findings. Treatment outcome was assessed at 16 weeks after randomisation by an independent clinical psychologist blinded to treatment allocation.

Both ABFT and TAU were provided for a minimum of 16 weeks but could be extended if deemed necessary by the therapist. ABFT consisted of weekly sessions as well as extra parent sessions in the early part of therapy. ABFT was delivered according to an available draft of the treatment manual (16). TAU was not manualised, and the therapists were free to provide the treatment they considered most appropriate. Adolescents allocated to ABFT on average received a higher number of sessions than those allocated to TAU, although some sessions in ABFT are conducted with parents alone. The mean number of sessions was 28.66 (SD 8.32) in ABFT, and 19.73 (SD 6.49) in TAU. When trial measurements or assessments indicated that adolescents were at high risk of self-harm or suicide, the therapist assigned to the case was immediately notified by study staff.

Clinicians were trained in ABFT for the purpose of the trial. Training consisted of a day-long introductory seminar, followed by a three-day workshop, as well as reading the treatment manual. Clinicians providing ABFT were required to have completed one case of ABFT under supervision before treating patients allocated to ABFT in the trial. All ABFT sessions were videotaped for supervision purposes. Weekly supervision by an experienced ABFT therapist was intended, but not achieved in practice. For the duration of the trial, the clinicians in the ABFT arm met nearly weekly and did peer supervision, and 42% of these sessions were also attended by a certified ABFT therapist and trainer. Clinicians in the TAU arm were also recruited from the regular staff of the CAMHS, and treated patients in the trial as part of their regular patient workload. None of the ABFT therapists treated patients allocated to TAU. Access to supervision for clinicians providing TAU varied by clinical experience, but all had access to discussing cases in multidisciplinary teams. Most of the clinicians across both treatment arms were trained as clinical psychologists (>75%), but there were also a small number of clinical social workers, clinical pedagogues and medical doctors in resident training providing treatment.

According to the protocol registered in clinicaltrials.gov (NCT01830088), our primary and secondary outcome measures were supposed to be collected at 12, 24 and 48 weeks after treatment start. A four week waiting period from randomisation to treatment start was planned, but this was not feasible due to the severity of the depression for many patients and Norwegian standards of care. Consequently, the treatment period was extended from 12 to 16 weeks, but the time from randomisation to outcome assessment was unchanged.

Diagnostic evaluations were conducted with the K-SADS-PL (29). Interrater reliability of the diagnosis of Major Depressive Disorder was established by an independent clinician conducting a blinded rescoring of a subsample of 28 videorecorded interviews, including both excluded and included patients. κ for current Major Depression was 0.56, indicating fair interrater reliability (30).

The primary outcome measure of the clinical trial was the total score on the GRID-HAMD, which is a version of the Hamilton Rating Scale for Depression that includes a structured interview guide, and scoring guidelines for weighing severity and frequency of symptoms to a composite score (27). The GRID-HAMD has previously shown excellent interrater reliability (31). Interrater reliability was assessed in the same way as for diagnoses, with the same subsample. The two-way mixed, consistency, average-measures intraclass correlation coefficient for the total score (32) was 0.89, indicating good interrater reliability.

Parent-adolescent conflict was measured with the Perception of the Dyad subscale of the Conflict Behavior Questionnaire (CBQ, 33). This scale consists of 16 items rated true or false concerning the current state of conflict in a parent-adolescent relationship, and was completed by parents as well as the adolescents separately for each parent. Example items are “My [parent/child] and I speak to each other only when we have to”, “At least once a day we get angry at each other” and “We have enjoyable talks at least once a day” (reverse scored). The CBQ was translated to Norwegian for this study, and a blind reverse translation was approved by the original author.

Analyses included all patients randomised to treatment regardless of adherence to study treatment or procedures, in accordance with intent-to-treat principles. One patient withdrew consent and was omitted from all analyses. We conducted analysis within a Bayesian modelling framework, with estimation by Hamiltonian Monte Carlo as implemented in the statistical modelling platform Stan, using the RStan package (version 2.19.2, 34) for R (version 3.6.1, 35). The results of a Bayesian analysis are distributions that show the probability of different model parameter values, conditional on the data and the model. For readers unfamiliar with Bayesian statistics, Baldwin and Larson (36) provide a very accessible introduction to the use of Bayesian linear regression in clinical psychology.

To improve the measurement precision of the hypothesised moderator variable we used a latent variable rather than raw scores, which we obtained by fitting a two-parameter logistic item response model to the CBQ data. The Stan platform is well suited for estimating IRT models, which can be embedded in a larger model of interest (37). Adolescents completed the CBQ separately for each parent, and each parent completed the CBQ for their relationship to the adolescent. In the majority of cases this gave four different ratings of the degree of parent-adolescent conflict, two by the adolescent for each parent, and one by each parent. We chose to model all four ratings as potential moderators, fitting these models separately. We used the report of all four informants to fit the item response model, specifying the four latent conflict variables to have a multivariate normal distribution, with means of 0 and standard deviations of 1, and constraining item parameters to be equal across informants. Checking this assumption of measurement invariance (38) resulted in removal of two items. Visual inspection of the posterior distributions of item characteristic curves plotted against the data indicated good fit for the remaining items. These plots as well as further details concerning checking of measurement invariance are available online at DOI: 10.17605/OSF.IO/KPJC6.

We specified a hierarchical linear regression model with pre- and post-treatment GRID-HAMD scores nested within patients as the outcome variable, and a random intercept for each patient. The model included terms for the predictor variables time, treatment allocation and parent-adolescent conflict, and interaction terms for treatment by time, conflict by time and a three-way interaction of treatment by conflict by time. The three-way interaction estimates the moderator effect of interest, while the conflict by time interaction estimates a non-specific predictor effect. Treatment allocation was coded as -0.5 for treatment as usual and +0.5 for ABFT, which allows the coefficients for treatment or interactions with treatment to be interpreted as the predicted difference between the treatment groups, with the sign indicating the direction of the difference (9). We standardised the GRID-HAMD scores by subtracting the median score across both time points of 21 and dividing by two times the median absolute deviation of 8, as standardisation can improve Hamiltonian Monte Carlo estimation and simplifies specification of reasonable priors (37).

Regression models with normally distributed errors are sensitive to outliers (39). Psychotherapy outcome is known to be influenced strongly by extratherapeutic factors (40), which increases the probability of having multivariate outliers. To avoid having outliers influencing slope estimates disproportionally to the bulk of observations we specified a Students’ t-distribution with five degrees of freedom to the errors, giving a robust estimation of regression coefficients (37, 41).

After fitting the model, we used exact leave-one-out cross-validation (LOO-CV), leaving out one patient at a time, to compare the model to three less complex models, repeating this across all four informant perspectives on conflict. LOO-CV estimates the expected log posterior density (ELPD), indicating how well the model is expected to fit new data from the same distribution (42). The first of these three models had the term for the moderator effect removed, making it a model with conflict as a non-specific predictor of outcome. The second had all terms involving conflict removed, making it a model of different effects of treatment allocation over time. The third had all terms involving both conflict and treatment removed, making it a model of the effect of time alone. For clarity, we will term these four models “Moderator”, “Non-specific predictor”, “Treatment” and “Time” when comparing them.

There was a substantial number of patients missing outcome data (38%). In most cases, outcome data was missing due to adolescents declining to come for assessment sessions or not showing up for such sessions, as well as lacking resources and administrative capacity to follow up further in such cases. Bayesian data analysis provides a natural way to handle such missing observations, by estimating these as unobserved parameters of the model, which ensures that the loss of information due to missing data is reflected in the posterior distribution as increased uncertainty of model parameters such as regression coefficients (43). When single items were missing from the CBQ, we estimated the latent variable of the IRT model from the observed items. In some cases, the entire CBQ was missing, and in these cases the latent variable was also estimated as a model parameter. This applied to 12% of adolescent reports of conflict with father, 28% of father reports of conflict, 3% of adolescent report of conflict with mother and 5% of mother reports of conflict. In some cases of missing reports of father-adolescent conflict, this was due to fathers not participating in treatment.

In a Bayesian data analysis, a prior distribution must be assigned to all model parameters, representing our prior knowledge of these parameters. This prior distribution is combined with the likelihood of the data to produce the posterior distribution of the parameters. This allows us to include information and assumptions on what ranges of a parameter are at all reasonable. When reading the results of a Bayesian data analysis, the reader should examine the prior distributions used and assess whether they are reasonable assumptions, and these should hence always be reported. The prior distributions used in this analysis and their justifications are summarised in Table 1.

We fitted all models using four chains with the default Stan algorithm, 1000 warmup iterations and drawing 3500 samples from each chain. Gelman-Rubin statistics (R′) were below 1.01 for all parameters. Other Stan convergence diagnostics also indicated convergence for all chains and valid sampling from the posterior.

To understand the implications of a fitted model, plotting its predictions across the range of a predictor variable can be helpful. Figure 2 shows the predicted change in GRID-HAMD score from baseline to outcome across the range of parent-adolescent conflict for all four informants, with the different lines representing the two treatment conditions (red for ABFT and black for treatment as usual). The uncertainty of the prediction is visualised by shading showing the 90% HDI. The points are the observations used to fit the model.

FIGURE 2.

Given the uncertainty in these estimates, they must be interpreted cautiously. For conflict with mother, the pattern is similar for both adolescent and mother report, with better outcomes predicted for ABFT relative to treatment as usual at high levels of conflict, and the opposite at low levels of conflict. This is not the case for conflict with father. For adolescent report the posterior distribution of the regression coefficient for a moderator effect has a mean of approximately 0, implying no moderation. For father report, the moderator model implies that parent-adolescent conflict is associated with worse or better outcomes in treatment as usual only, but that outcome in ABFT does not vary over father-reported conflict.

Table 5 shows the differences in ELPD (expected log posterior density, obtained by LOO-CV) between the four models that were compared. It should be noted that the standard errors of these differences are known to be optimistic, especially in small samples, and a difference of four standard errors or more has been recommended for selection of one model over another (42, 46).

Cross-validation clearly shows that the model “Time” has a better expected out-of-sample fit than “Treatment”, with a difference in ELPD larger than ten times the standard error. The model “Time” also fits better than the model “Non-specific predictor” across all four informants, with differences in ELPD of more than five standard errors. The picture is less clear for the comparison of the “Moderator” model to “Time”. For adolescent report of conflict with father “Time” is clearly better, with a difference larger than ten times the standard error. For father-report of conflict, and for adolescent and mother report of conflict with mother, the differences are too small relative to their standard errors to support selecting either model over the other.