Clinical problem

Depression is a major contributor to the global burden of disease and is projected to be the leading cause of disability in high-income countries by 2030.¹⁶¹ There is good evidence that CBT is an effective treatment in previously untreated episodes of depression.^162,163 However, there is considerable variation in patient response to CBT, with significant proportions responding either not at all or only partially. It is therefore clinically useful to identify which patients will respond well to which treatments. Reliable evidence informing this issue remains elusive, however, and clinicians often decide which patients to refer to which treatment, based on implicit beliefs about patient suitability.¹⁶⁴ In line with the current drive towards stratified medicine that aims to target interventions at subgroups of patients who are likely to respond,^165,166 research is needed to identify reliable moderators or effect modifiers of treatment response.

It is important to distinguish between predictors and moderators. Predictors are prognostic factors associated with disease outcomes irrespective of treatment, whereas moderators or effect modifiers are associated with different treatment responses.¹³⁹ In other words, a moderator will lead to a smaller or larger difference between active and comparator groups. Understanding of potential moderators is clinically useful as this would enable clinicians to base treatment choices on the individual’s likelihood of benefiting from a given treatment. A variable is established as a moderator by testing for interactions between that variable and two or more treatment options, ideally within the context of a controlled trial.¹⁶⁷ Studies designed specifically to test for interactions are large, expensive, and therefore rare. Using existing data from good-quality and well-controlled clinical trials is an efficient and cost-effective alternative.¹⁶⁸ Although such secondary analyses suffer from low statistical power, they are prone to false-positive findings owing to multiple testing. Caution is therefore required when interpreting findings from a single study. Consistent findings across studies are required before we can consider moderators as clinically informative and, ideally, the field should aim for meta-analyses of randomised studies using individual patient data to achieve sufficient statistical power.¹⁶⁹

To date, studies reporting moderators of response to CBT in controlled trials have used small sample sizes, randomising fewer than 63 patients per CBT arm,^170–174 or have compared CBT to antidepressant treatment,^172,173 or have focused on adolescent or elderly populations.^175–177 With such small sample sizes these studies were almost certainly underpowered,¹⁶⁷ and although understanding which of two treatment options is likely to produce the best outcomes is important, antidepressants and CBT are often prescribed together in practice. In most health services, antidepressants are widely available and access to psychotherapy is limited. CBT is often reserved for those patients who have not responded to antidepressant medication. We have previously reported effect modifiers for online CBT as an adjunct to usual care, which included the option of antidepressants where prescribed, and just over half the sample were taking antidepressants at baseline.¹⁶⁸ However, to date, no research has examined moderators of response to CBT as a ‘next-step’ treatment for primary care patients who have not responded to antidepressants. Identifying reliable moderator variables in this population will inform treatment options for depressed patients who do not respond to antidepressants. To date, the existing literature in this area is based on studies of CBT in populations without TRD.

Demographic factors

A few studies have examined whether or not demographic variables such as age, gender, education and marital status are moderators of response to CBT in adults using appropriate tests for interaction in controlled studies.^168,172,174 In a stepwise regression, Fournier and colleagues¹⁷² found that being married, unemployed or having more antecedent life events were associated with better response to CBT than antidepressants. However, patients often receive CBT as an adjunct treatment in addition to antidepressant medication. In a previous study¹⁶⁸ with a sample twice the size of that reported by Fournier and colleagues,¹⁷² we found that being separated, widowed or divorced or having fewer recent stressful life events were associated with better responses to CBT compared with a waiting list control. In our study, both arms received care as usual, including antidepressants if prescribed by the GP,¹⁶⁸ which may account for the discrepant findings. Furthermore, in contrast with Fournier et al. ¹⁷² we divided our ‘unmarried’ participants, into ‘single’ or ‘separated/widowed/divorced’.¹⁶⁸ Educational attainment and age were not found to modify response to CBT.^168,172

Illness characteristics

To date, pre-treatment severity of depression is the most reliable moderator of response to CBT, with the more severely depressed benefiting most.^{168,177–179} The evidence suggests that mild depression seems to recover well irrespective of treatment, whereas severe depression gains most from CBT.¹⁶⁸ However, meta-analytic findings that rely on aggregate data^179,178 and issues of scaling confuse these severity findings, which may be an artefact of assessing outcomes using continuous measures. For example, a 5-point reduction in scores for someone whose baseline score was 50 is a proportionally smaller improvement than 5 points for someone with a baseline score of 15.

There is no evidence that history, chronicity and type of depression are moderators of CBT response.^168,172 The literature on comorbidity is also mixed. Patients in the STAR*D study with anxious depression who were partial or non-responders to citalopram responded less well to either CBT or an alternative antidepressant as a second-line treatment but there was no evidence for effect modification.¹⁷¹ Asarnow et al. ¹⁸⁰ identified comorbidity, including anxiety, as a moderator, with increased comorbidity associated with increased response to combined CBT and antidepressants relative to antidepressants alone. However, anxiety did not modify response to CBT in a study of depressed adolescents.¹⁷⁵

Personality traits, cognitions and psychological mindedness

Assessing individuals’ suitability for therapy is an important part of clinical practice, which often focuses on interpersonal skills, personality and psychological mindedness.¹⁶⁴ The lack of clear evidence of effect modification from appropriately controlled studies, however, illustrates that this practice is not evidence based. Indeed, few studies have investigated these variables as moderators. To date, there is no evidence that the personality trait neuroticism is a moderator.^172,181 Patients with lower dysfunctional attitudes have been found to do better in treatment arms (CBT and antidepressants) relative to pill placebo,¹⁷⁴ whereas other studies found no evidence for effect modification.^172,176 Clinicians believe it is important to individualise treatment in line with particular patient presentations,¹⁸² and it seems likely that cognitive-based therapies will be most effective in those with high levels of psychological awareness. However, no studies have directly investigated metacognitive awareness as a moderator in an appropriately controlled trial.

Aims

The aim of the present analysis was to examine potential moderators of response to CBT given as an adjunct to usual care that included pharmacotherapy as a ‘next-step’ treatment for patients whose depression had not responded to treatment with antidepressants using data from the CoBalT trial. By examining moderators in this group, we aim to inform decisions of whether or not to refer such patients for CBT. We examined the modifying effects of demographic, life events, illness, comorbidity, personality traits and cognitive variables.

Participants

This was a secondary analysis of data collected as part of the CoBalT trial, which has been described in detail earlier (see Chapter 2 ). Brief details are outlined below.

Individuals were eligible for the trial if they were aged between 18 and 75 years, were currently taking antidepressant medication and had been doing so at an adequate dose for at least 6 weeks, scored ≥ 14 on the BDI-II⁴⁶ and met the ICD-10 criteria for depression (assessed using the CIS-R^47,183). Participants were randomised to one of two groups: (1) usual care or (2) CBT in addition to usual care. Treatment allocation was stratified by recruitment centre and minimised by baseline BDI-II score, whether participant’s general practice had a counsellor (yes/no), prior treatment with antidepressants (yes/no) and duration of their current episode of depression (< 1 year; 1–2 years; ≥ 2 years) in order to achieve balance in these important (design) variables across the treatment arms. Participants were followed up at intervals of 3 months for 1 year, with the BDI-II being completed at baseline, 6 and 12 months.

Outcome

The outcome variable used in this secondary analysis was BDI-II score treated as a continuous variable at 6 and 12 months’ follow-up analysed as a repeated measure. We treated BDI-II score as a continuous variable in this exploratory study to retain maximum power and ensure comparability of findings with previous studies of moderation.^168,172 This is in contrast with the main trial, where the primary outcome was a binary variable representing a reduction in BDI-II score of at least 50% compared with baseline (see Chapter 2 , Primary outcome).

Moderators

All data on potential moderators were collected as part of the baseline assessment, prior to randomisation. The potential moderators were grouped into three general classes: (1) demographic and life factors; (2) illness characteristics; and (3) personality, cognition and psychological mindedness.

Statistical analysis

Treatment effect was defined as the (adjusted) difference in mean BDI-II outcome score (as a continuous variable) between the usual care and intervention arms. Separate repeated measures linear regression models were carried out for each potential moderator. The model included an interaction term between the moderator and treatment allocation, and adjusted for the design variables (including baseline BDI-II score) and time. Further models, containing a three-way interaction (moderator by treatment allocation by time) were carried out to investigate whether or not effect modification varied across time. Repeated measures regression models were also stratified by each level of the potential moderators to illustrate any interaction effects.

Baseline characteristics

As reported earlier (see Table 19 ), the randomised groups were similar in terms of the stratification and minimisation variables (baseline BDI-II score, whether participant’s general practice had a counsellor, prior treatment with antidepressants, and duration of their current episode of depression), age, gender and demographic factors. The two groups were also similar in terms of the other potential treatment moderators investigated (Table 81) that were not reported earlier.

TABLE 81

Comparison of additional baseline characteristics between randomised groups

Adherence to the intervention

The level of adherence to the intervention (defined as the mean number of CBT sessions attended) were generally very similar across the levels of the potential moderators investigated (Table 82).

TABLE 82

Adherence to CBT intervention

Effect modification by potential moderators

The results obtained from the repeated measures regression models suggested that age was the only variable for which there was any evidence of an interaction between a potential moderator and the intervention, implying that age may modify the effectiveness of CBT. The interaction coefficients became more negative the higher the age category, suggesting that the higher the age category the greater the benefit of treatment (p-value for interaction effect = 0.012; Table 83 ). When age was used as a continuous variable the conclusion was the same, evidence for greater treatment derived benefit the older the subject (interaction coefficient = –0.20, 95% CI = –0.37 to –0.02, p = 0.027).

TABLE 83

Results from repeated measures regression models testing variables for potential modification of treatment effect

The regression analyses were also carried out separately at each level of the potential moderator variables in order to illustrate the findings. The adjusted differences in mean BDI-II scores between the levels of the investigated variables were similar, had overlapping CIs and did not show any clear trends except for age, in which older individuals had a larger treatment response (see Tables 84 and 85 ). The three-way treatment × moderator × time interactions suggested that there was no evidence that the relationships between any of the investigated potential moderators and the intervention varied over time (Table 83).

TABLE 84

Adjusted differences in mean BDI-II score between randomised groups to further illustrate any interaction effects

TABLE 85

Results from repeated measures regression models when carried out separately for each level of the potential moderator

Summary of main findings

This is the first study testing for moderators of response to CBT as a ‘next-step’ treatment for primary care patients who have not responded to antidepressants. Of the 14 variables assessed, age was the only variable with some statistical evidence for effect modification, with older patients benefiting the most from CBT. We found no evidence of effect modification by any other demographic, life, illness, personality trait or cognitive variable. Insufficient power prevents conclusive interpretation of such null findings. However, our findings suggest that it would be premature to adopt a stratified approach to prescribing CBT as a ‘next-step’ treatment for individuals who have not responded to antidepressants.

Strengths and limitations

The limitations associated with post hoc subgroup analyses should be borne in mind when interpreting our findings.¹⁶⁷ Although the sample size (n = 469) is one of the largest RCTs of CBT to date, it is small for testing interactions, creating uncertainty about the reliability of the estimates. In addition, multiple testing increases the likelihood of chance findings. However, we tested 14 different variables and found evidence for only one moderator. Given the move towards stratified medicine,¹⁶⁵ it is important to discern for which patients CBT is likely to work and for what reasons. Pragmatically, CBT is often reserved for the patients who have not responded to antidepressants. Hence, identifying moderators of CBT response in this population is clearly clinically important, and ours is the first large-scale controlled study to examine effect modifiers of CBT offered as a ‘next-step’ treatment for non-responders to pharmacotherapy.

Demographic and life factors

Age was the only variable with evidence for effect modification, with older patients benefiting the most from CBT. There is no precedent for age as a moderator in previous studies of CBT in populations without TRD,^168,172 so we treat our result with caution, and it may be a type I error. In contrast with RCTs of previously untreated episodes of depression, the mean age of patients in CoBalT was higher, with over half of the sample being ≥ 50 years when they entered the study. We would not expect this in itself to influence the findings in terms of the pattern of coefficients, especially given the good balance between the trial arms with respect to age. Yet it may have increased our power to detect this particular interaction compared with other studies with a younger age distribution. Alternatively, it may reflect something specific to the treatment-resistant population. CBT was most effective for patients over 40 years, and least effective in patients aged 30–39 years. It is unclear why CBT was not beneficial in this younger subgroup, but it is worth noting that given the small numbers (n = 61) the CIs around the estimate are wide, providing no evidence for either treatment benefit or harm. Further research is required to assess whether this finding is replicated.

In contrast with previous research,^168,172,174 we did not identify marital status or stressful life events as moderators. The point estimates for marital status were consistent with single individuals gaining least from CBT, but there was no statistical evidence for effect modification (p = 0.34). The estimates for life events showed no evidence or even a suggestion of support for previous findings in non-TRD samples.^168,172

Illness characteristics and comorbidity

In contrast with previous studies,^{168,177–179} pre-treatment severity of depression did not moderate response to CBT. This may reflect the nature of our treatment-resistant sample; in CoBalT patients were selected for their non-response to antidepressants. In our previous study, mild depression seemed self-limiting, improving equally well irrespective of receiving CBT or waiting list control. By contrast, CBT was particularly effective for severe depression, which did not improve in the waiting list arm. By definition, the depression in patients recruited to CoBalT was not self-limiting as we selected patients through the resistance of their symptoms to pharmacotherapy. This may explain the absence of effect modification by severity in this group. It is of note, however, that baseline severity in CoBalT was similar to other RCTs of depression in the UK.^109,185,186 The CoBalT sample was nevertheless more ill in terms of chronicity, number of previous episodes, comorbidities and non-response to medication. This suggests that to capture the extent of illness that we see clinically then we need to account for both severity and chronicity, especially in those whose symptoms are resistant to antidepressants. However, an a priori subgroup analysis conducted in the main trial found no evidence that the degree of treatment-resistance modified response to CBT (see Chapter 3, Subgroup analyses).

Personality, cognitions and psychological mindedness

The literature on assessing individuals’ suitability for CBT is based more on clinical opinion¹⁶⁴ than empirical evidence, and often focuses on interpersonal skills (which we were unable to investigate), personality and psychological mindedness. Consistent with previous research assessing personality traits in untreated episodes of depression,^172,181 we found no evidence that neuroticism was a moderator of response to CBT in treatment-resistant individuals. Furthermore, dysfunctional attitudes and metacognitive awareness were not associated with differential response. The practice of selecting patients for CBT based on assessments of personality and psychological dysfunction and awareness remains empirically unsupported. Indeed, such practice might be harmful, preventing individuals, who might actually benefit, from receiving CBT.

Clinical implications

Cognitive behavioural therapy as an adjunct to usual care is an effective ‘next-step’ treatment for patients whose depression has not responded to treatment with antidepressants. This is an important finding, as, in practice, limited availability means that referral for CBT often only follows non-response to first-line treatment with antidepressants. However, to further improve patients outcomes by tailoring treatment in line with stratified medicine¹⁶⁵ it is helpful to understand if there are any factors associated with differential treatment response. We found that response to CBT differed with age, with older age groups benefiting more. Given the small numbers of patients and wide CIs in this subgroup, we caution against using age to inform treatment decisions until further research replicates this effect. We found no evidence to suggest that non-response varied systematically with other patient characteristics. Until research replicates the age finding, and in the absence of other clear and reliable moderators, consideration should be given to offering CBT to all individuals where antidepressant medication has failed. Not all patients will respond, but, as we have only preliminary evidence as to whom these might be, consideration should be given to offering CBT to all patients in this severely ill group.