Background

The global burden of Major Depression for the society but also for the individual quality of life is extremely high. Depressive disorders are discussed to be amongst the first reasons for chronic conditions for mental disorders in Europe. Eleven % of all years lived with disability are due to depressive disorders, as assessed by the WHO (Global Health Estimates, 2016). Psychological treatments can significantly reduce symptoms of Major Depression and, therefore, are recommended in the internationally acknowledged guidelines (APA, 2019; National Collaborating Centre for Mental Health, 2009, last update 2018) and national guidelines (DGPPN et␣al., 2015). Many studies provide support for the efficacy of cognitive-behavioural therapy (CBT) for several mental disorders, especially for depressive disorders (Butler, Chapman, Forman, & Beck, 2006; Cuijpers, Cristea, Karyotaki, Reijnders, & Huibers, 2016; Hofmann, Asnaani, Vonk, Sawyer, & Fang, 2012). A few RCTs addressing the efficacy of hypnotherapy (HT) for depressive disorders have been conducted (Alladin & Alibhai, 2007; Butler et␣al., 2008; Chiu, Lee, & Lam, 2018). Positive results for HT were found. For example, HT was superior to care as usual (Chiu et␣al., 2018), to a psychoeducational control group but without reaching significancy (Butler et␣al., 2008), and a combined cognitive hypnotherapeutic treatment performed better than cognitive therapy alone (Alladin & Alibhai, 2007). However, previous RCTs had small sample sizes or methodological quality was low. In a recent RCT, it was found that HT was not inferior to CBT in the treatment of major depression (Fuhr et␣al., 2021). Even if different forms of psychotherapy are available, there is still a large number of patients without adequate response to the treatment. Several studies aimed to identify baseline characteristics of patients that are valuable in the prediction of later treatment outcome. For example, the number of previous episodes and the symptom severity at baseline are negatively associated with treatment response (Hamilton & Dobson, 2002; Hoberman, Lewinsohn, & Tilson, 1988). Even with several sociodemographic, clinical, cognitive and personality variables that were assessed before treatment, only around 20% of variance could be explained concerning treatment outcome (Carter et␣al., 2011). Patients who are likely to benefit from a particular treatment should be identified at an early stage to reduce time and direct and indirect costs. In the last years, many studies were conducted to identify different symptom courses of patients and their relationship to the treatment outcome (Beard & Delgadillo, 2019). The majority of these studies focussed on Major Depression as the primary disorder mostly with a pharmacological, mixed or psychotherapeutic treatment (Arnow et␣al., 2007; Gildengers et␣al., 2005; Lutz, Stulz, & Köck, 2009). They found that patients who respond to the treatment in the first weeks (early responders) were more likely to show a better outcome at the end of the treatment (Braun, Strunk, Sasso, & Cooper, 2015; Gilboa-Schechtman & Shahar, 2006; Steidtmann et␣al., 2013; Tadić et␣al., 2010). Little studies focussed on treatment with psychotherapy alone (Braun et␣al., 2015; Steidtmann et␣al., 2013).

According to the phase model of psychotherapy (Howard, Lueger, Maling, & Martinovich, 1993), three phases can be distinguished in psychotherapy: The change in subjectively experienced well-being in the remoralization phase, the reduction in symptomatology in the remediation phase and the enhancement in life-functioning in the rehabilitation phase. The change after the first two sessions, for example, could be allocated in the remoralization phase, where the first actions with the therapist are able to enhance the patients’ well-being, even before some formal treatment sessions had happened. The change in this remoralization phase leads to a subsequent (rapid) symptom improvement during the first four sessions in the remediation phase. In the last phase after session five onward there Howard et␣al. (1993) found a slow but constant improvement not only in the subjective well-being and the symptoms but also in the global life functioning.

Results on early change seemed to confirm the model of psychotherapy. For example, patients with an early improvement of 20% or more in the first two weeks showed a stable response up to eight weeks later, not only in pharmacological treatment but also in CBT (Tadić et␣al., 2010), and in a psychological or combined treatment (Van et␣al., 2008). Gilboa-Schechtman and Shahar (2006) specifically referred to Howard’s model of psychotherapy in their research article. They found that patients who showed a rapid relief in distress during the first four weeks were more likely to have lower depressive symptoms after 16 weeks of treatment and at follow-ups. Nevertheless, their definition of early change included four instead of two weeks as in Howard et␣al. (1993), and thus, can be allocated in the combined remoralization and remediation phase of the model. Treatment was antidepressant medication (Gilboa-Schechtman & Shahar, 2006). Rush, Kovacs, Beck, Weissenburger, and Hollon (1981) were amongst the first of pointing to an early improvement phenomenon and found up to 60% of symptom improvement during the first four weeks in a treatment of eleven weeks with either cognitive therapy or antidepressant medication (Rush et␣al., 1981). In contrast, patients without early improvement in the first weeks were more likely to show treatment non-response (Steidtmann et␣al., 2013; Tadić et␣al., 2010) or higher symptom severity at the end or follow-ups (Gilboa-Schechtman & Shahar, 2006).

As a first interim conclusion, the number of weeks accounting for early change varied across studies between 2 and 10 weeks (Busch, Kanter, Landes, and Kohlenberg 2006; Lewis, Simons, & Kim, 2012; Lutz et␣al., 2009; Tadić et␣al., 2010; Van et␣al., 2008). This complicates integrating results to models of psychotherapy like the one of Howard et␣al. (1993).

Furthermore, the studies on early improvement also differ regarding the methods used for assessing the overall effect of a treatment. Most studies used the symptom severity as assessed with self-reported or clinical-administered questionnaires at the end of a treatment (Gilboa-Schechtman & Shahar, 2006). Other studies used categorical data such as response or remission rates after treatment or at follow-ups (Steidtmann et␣al., 2013; Tadić et␣al., 2010). Rubel, Rosenbaum, and Lutz (2017) criticized that most previous studies used an average effectivity to measure the outcome of psychotherapy. They used multilevel modelling and described advantages of using the individual within- and between-patient variances to predict the associations between specific actions in psychotherapy and the symptom change in a sample with 409 patients participating in a web-based intervention (Rubel et␣al., 2017). Following their approach, it would be instead interesting to focus on individual change of symptoms, which can be also done on a session-to-session level rather than on average or percentage symptom improvement. Session-to-session changes and routine outcome monitoring offer the possibility to assess individual symptom changes and related actions during treatment (Lambert, Whipple, & Kleinstäuber, 2018; Rubel et␣al., 2017). The advantages of session-to-session analyses have been also referred by Strunk, Brotman, and DeRubeis (2010). For example, they allow for short changes, missings at random, and a more precise estimation of what is happening during treatment instead of using predictions over the whole course of the treatment (Strunk et␣al., 2010). Also, from a clinical perspective, it would be important to identify patients at an early stage of psychotherapy that would benefit from further treatment or not. Some researchers concluded that clinical decisions about patients’ treatment should be conducted based on their early symptom change (Gilboa-Schechtman & Shahar, 2006; Percevic, Lambert, & Kordy, 2006). Patients with early non-response should be treated differently since they showed severe symptoms at follow-ups (Gilboa-Schechtman & Shahar, 2006). But it still remains unclear, if patients without early change are less likely to show a change of symptoms later in treatment than patients with early change. This could not be answered sufficiently by previous research since the outcome was measured in most studies at the end of the treatment and not on a weekly or session-to-session base. Thus, we still do not know, whether early symptom change predicts the subsequent course of treatment and, therefore, might be considered as the ‘temporal precedent’ (Feeley, DeRubeis, & Gelfand, 1999) of later symptom change. Percevic et␣al. (2006), for example, investigated different trajectories of symptomatic changes during a treatment with psychotherapy on a session to session level. They did not find that early improvement predicted later changes in treatment. As already stated by Percevic et␣al. (2006), there was also a temporal overlap between early response and overall response as it was investigated before: In other words, the change during the first weeks (used for the definition of the early response) was included in the overall symptom change to the end of the treatment (defined as overall) response in most of the previous research (Beard & Delgadillo, 2019; Lewis et␣al., 2012; Tadić et␣al., 2010).

Aims of the study

The present study aimed to investigate whether the individual symptom course in the first weeks of the treatment predicts the subsequent individual symptom course on a session-to-session level. Due to the high methodological heterogeneity in previous studies, we wanted to compare different models where we vary the number of weeks for defining early change. We tested if the symptom changes after session four, five, six and seven onward on a session-to-session-level could be predicted by the early symptom change during the first two, three, four and five weeks in a psychotherapy trial with overall 20 weekly sessions.

Methods

The present study is part of a randomized controlled trial (RCT), which focussed on the treatment of Major Depression with psychotherapy. It was found that Hypnotherapy (HT) was non-inferior to CBT in reducing depressive symptoms in patients with mild-to-moderate depressive episodes (Fuhr et␣al., 2021). The present study is of exploratory nature and was not planned before. We calculated all models testing early change for both intervention groups separately (see Table S1 for a detailed overview). Analyses in both groups lead to a comparable conclusion regarding the prediction of the subsequent symptom course. Therefore, we decided to omit the treatment condition as predictor variable in the main paper and focussed on a general model of symptom change over time for all patients in the trial, and thereby also increase the power of the models.

Trial design

In the RCT, 152 participants were randomized in a ratio of 1:1 to either 20 individual sessions with HT or CBT. Both treatments covered up to 20 sessions and were parallelized in length. Depressive symptoms were assessed weekly. The trial followed the CONSORT and GCP guidelines and was conducted in accordance with the Declaration of Helsinki. Ethical approval was given by the Ethics Committee of the Medical Department of the study site (061/2015BO2). The trial was registered within ClinicalTrials.gov (NCT02375308). The study protocol was already published elsewhere (Fuhr, Schweizer, Meisner, & Batra, 2017).

Participants

We recruited and screened the participants for our study between May 2015 and December 2016 at the University Hospital of Psychiatry and Psychotherapy at the study site in Germany. Subjects gave their informed written consent after the procedures and possible side effects were fully explained. Patients with a current mild to moderate episode of Major Depression (MD) according to the Diagnostic and Statistical Manual of Mental Disorders—Fifth Edition (American Psychiatric Association, 2013) were included in the study. The clinical interviews were done by seven different ratters with a degree in psychology (at least bachelor), a university course in conducting clinical interviews, and a half-day training. In cases of antidepressant medication, a stable medication in the last three months without planned changes during the therapy period was required. Patients with the lifetime diagnosis of a bipolar disorder or psychotic disorder, the diagnosis of chronic MD, actual high suicidality or other severe primary psychological disorders (active alcohol or drug dependence, actual post-traumatic stress disorder or anorexia nervosa), or patients receiving outpatient psychotherapy within the last twelve months were excluded from the trial.

Interventions

The individual psychological treatments in both conditions included 20 sessions during a period of six months. The treatments were based on German treatment manuals of CBT (Hautzinger, 2013) and HT (Wilhelm-Goessling, Schweizer, Duerr, Fuhr, & Revenstorf, 2020). On average, 19 weekly sessions were attended. Four therapists per treatment condition treated 12–24 patients over the study period at the study centre.

Patient Health Questionnaire—–depression module (PHQ-9)

Depressive symptoms were assessed via the self-rated PHQ-9 before treatment and on a weekly basis during the therapy period with nine items regarding the past seven days (Kroenke, Spitzer, & Williams, 2001). Range of the sum score is 0–27. Cronbach’s Alpha of the PHQ at baseline (before session 1) in the current sample was α = .68, compared to α = .86 to .89 in the original study (Kroenke et␣al., 2001). The low reliability could be due to the fact that we excluded patients from the trial when reporting suicidal ideas (which is one item of the PHQ-9).

Statistical analysis

This study used linear and quadratic mixed model (subsumed as LMM) analyses with random effects for each patient. In order to test the hypotheses, model comparisons via likelihood-ratio-tests (LRTs) were conducted. Furthermore, the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) were assessed. Both criteria assess how much information is lost compared to a model perfectly predicting the data, whilst also accounting for the number of parameters in a given model. Smaller values indicate a more adequate model, whereby differences between models smaller than 2 should not be interpreted.

Participants who did not fill out a questionnaire at session one (n = 3) were excluded from the analysis resulting in a sample of 138 patients after removing dropouts during the trial attending less than 16 sessions of treatment. Considering the lack of a uniform time period for early change, we computed separate analyses with four different endpoints at session three, four, five and six. Accordingly, early symptom change was defined as the individual percentage difference in the PHQ-9 score between session one (starting point) and three, four, five or six (endpoint), respectively, using the following formulae:

In case of a missing questionnaire at the endpoint, participants were excluded from this specific analysis (e.g. participants missing a questionnaire at session three would be excluded only from Analysis 1). This resulted in a dataset of 125 participants for Analysis 1, 106 participants for Analysis 2, 115 participants for Analysis 3 and 124 for Analysis 4. To achieve the intended separation of predicting and predicted variable, in all analyses, we modelled the course of depressive symptoms from the session following the endpoint onward (e.g. Analysis 1 modelled the symptom course from session four onward). In the models, we, therefore, recoded the session variable by subtracting 4, 5, 6 or 7 from the original session number.

In a first step, the best fitting LMM using only linear and quadratic time effects as fixed effects was selected. Subsequently, we assessed the necessity of random linear (and if applicable quadratic) slopes for each participant. Models used in this first step are labelled tLMM. In the second major step, we tested whether the addition of early change as a metric variable significantly improved the model fit. Models used in the assessment of early change are called eLMM. Finally, a brief oversight over the selected model’s parameter estimation is given. All analyses were computed in R (R Core Team, 2021) using the lme4-package (Bates, Maechler, Bolker, & Walker, 2015). We also calculated all models testing early change for both intervention groups separately (see Table S1).

A graphical inspection of the sample’s residuals indicates that the assumption of normality approximated the empirical distribution in an adequate way.

Power analysis

Since our analyses were of explorative nature, we conducted a post hoc simulated power analysis to assess whether our data suffice to detect a potential effect of early change on the further course of symptoms. In this analysis, we simulated several thousand data sets where the further course of symptoms depends on early symptom change, to a different degree. We then proceeded to fit the same models as in our main analysis, including all fixed and random effects. Subsequently, we calculated respective LRTs, assessing the proportion of significant test results regarding the interaction effect between early change and time. Our simulations showed that power in our sample is high enough to detect even an interaction effect of a small magnitude. For example, an increase in weekly PHQ-9 score reduction of 0.02 points per 10% additional early reduction could be detected with a power of (1 − β) = .90.

Results Sample

Of the total 152 randomized patients, 11 were excluded because they dropped out during the treatment before attending session 16, and three were excluded because of missing data in the PHQ-9 score of session 1. In total, 138 patients were included in the current study with either HT (n = 65) or CBT (n = 73). Mean age was 39.22 (SD = 14.49) and 70% of the participant were female. Sample characteristics of the overall sample are displayed in Table 1.

Table 1. Characteristics of the trial sample Variables Total (n = 138) M (SD) Age 39.22 (14.49) Depressive symptoms PHQ-9 at baseline 14.40 (4.23), range 5–23 Number of sessions 19.18 (1.33) n (%) Group HT 65 (47.1) CBT 73 (52.9) Sex (female) 96 (69.6) Highest educational level High school or higher 105 (76.1) No high school degree 33 (23.9) Antidepressant medication (yes) 52 (37.7) MDD subtype Recurrent 108 (78.3) Single episode 30 (21.7) Comorbidity (yes) 62 (44.9) Note CBT = Cognitive-behavioural therapy; HT = Hypnotherapy; M = mean; MDD = Major Depressive Disorder; N = number of patients; PHQ-9 = Patient Health Questionnaire – depression module; SD␣= standard deviation. Early and overall change

Mean early change over all patients was a symptom reduction of 2.69% (SD = 41.6) at session 3, 13.3% (SD = 32.1) at session 4, 13.8% (SD = 41.1) at session 5 and 16.3% (SD = 44.2) at session 6. Comparisons of the PHQ-9 scores from first and last therapy session showed, that 54.5% of patients responded to therapy (symptom reduction ≥ 50%) and 44.7% of patients remitted (PHQ-9 score < 5).

Time model

In analyses 2–4, LRTs selected the model with only a linear time effect (tLMM1) as the best model for the overall symptom course of individual PHQ-9 scores from after the endpoint onward. The BIC also supports this decision, whereas the AIC makes no distinction between the models with and without quadratic time effects (tLMM2 vs. tLMM1). As two of our criteria oppose the addition of quadratic time effects and the third one shows no preference, we selected tLMM1 as most appropriate. Only Analysis 1 selected an additional quadratic time effect, which is also supported by the AIC, whilst the BIC makes no distinction. Hence, we selected tLMM2 as most appropriate model for Analysis 1.

Furthermore, in all models, random intercepts and random linear slopes (tLMM3/tLMM4) are needed to appropriately describe the data. Both information criteria strongly support this claim (see Table 2).

Table 2. Model equations with corresponding AIC, BIC and LRTs Model Equation Tested parameter χ2 df p AIC BIC a. Analysis 1 tLMM0 yij = β0 + υ0i + εij 8,915.2 8,931.4 tLMM1 yij = β0 + β1tij + υ0i + εij β1 233.20 1 <.001 8,684.0 8,705.6 tLMM2 yij = β0 + β1tij + β2(tij)2 + υ0i + εij β2 18.01 1 .005 8,678.0 8,705.0 tLMM2 yij = β0 + β1tij + β2(tij)2 + υ0i + εij 8,678.0 8,705.0 tLMM4 yij = β0 + β1tij + β2(tij)2 + υ0i + υ1itij + εij υ1i 150.41 2 <.001 8,531.6 8,569.4 tLMM5 yij = β0 + β1tij + β2(tij)2 + υ0i + υ1itij + υ2i(tij)2 +εij υ2i <0.01 3 >.999 8,539.9 8,593.9 Analysis 2 tLMM0 yij = β0 + υ0i + εij 6,883.0 6,898.5 tLMM1 yij = β0 + β1tij + υ0i + εij β1 199.80 1 <.001 6,685.2 6,705.9 tLMM2 yij = β0 + β1tij + β2(tij)2 + υ0i + εij β2 1.86 1 .172 6,685.4 6,711.2 tLMM1 yij = β0 + β1tij + υ0i + εij 6,685.2 6,705.9 tLMM3 yij = β0 + β1tij + υ0i + υ1itij + εij υ1i 81.805 2 <.001 6,607.4 6,638.4 Analysis 3 tLMM0 yij = β0 + υ0i + εij 7,050.7 7,066.2 tLMM1 yij = β0 + β1tij + υ0i + εij β1 129.48 1 <.001 6,923.2 6,943.9 tLMM2 yij = β0 + β1tij + β2(tij)2 + υ0i + εij β2 2.99 1 .084 6,922.2 6,948.1 tLMM1 yij = β0 + β1tij + υ0i + εij 6,923.2 6,943.9 tLMM3 yij = β0 + β1tij + υ0i + υ1itij + εij υ1i 129.94 2 <.001 6,797.3 6,828.4 Analysis 4 tLMM0 yij = β0 + υ0i + εij 6,849.1 6,864.6 tLMM1 yij = β0 + β1tij + υ0i + εij β1 103.43 1 <.001 6,747.7 6,768.4 tLMM2 yij = β0 + β1tij + β2(tij)2 + υ0i + εij β2 2.77 1 .096 6,746.9 6,772.8 tLMM1 yij = β0 + β1tij + υ0i + εij 6,747.7 6,768.4 tLMM3 yij = β0 + β1tij + υ0i + υ1itij + εij υ1i 81.16 2 <.001 6,670.5 6,701.5 b. Analysis 1 tLMM4 yij = β0 + β1tij + β2(tij)2 + υ0i + υ1itij + εij 8,531.6 8,569.4 eLMM3 yij = β0 + β1tij + β2(tij)2 + β3ei + υ0i + υ1itij + εij β3 0.82 1 .366 8,532.8 8,576.0 eLMM4 yij = β0 + β1tij + β2(tij)2 + β3ei + β4(tij × ei) + υ0i + υ1itij + εij β4 1.56 1 .212 8,533.2 8,581.8 eLMM5 yij = β0 + β1tij + β2(tij)2 + β3ei + β4(tij × ei) + β5([tij]2 × ei) + υ0i + υ1itij + εij β5 1.15 1 .285