The analysis and reporting of this trial was undertaken in accordance with CONSORT guidelines.60,69,70
All statistical analysis was undertaken in Stata 11.2 (StataCorp LP, College Station, TX, USA), following a
pre-defined analysis plan agreed with the Trial Steering Committee (TSC). The primary comparative
analyses between the randomised groups were conducted on an intention-to-treat (ITT) basis without imputation of missing data.
Preliminary analyses
Descriptive statistics of the key sociodemographic and clinical variables were used to ascertain any marked imbalances at baseline, and to inform any additional adjustment of the primary and secondary analyses as appropriate.
Primary analysis
The primary analysis used logistic regression to compare the groups as randomised in terms of the primary
(binary) BDI-II outcome at 6 months, adjusting for stratification and minimisation variables (the ‘design
variables’: centre, baseline BDI-II score, access to a counsellor, prior treatment with antidepressants and
TABLE 4 Original and revised power calculations
Sample size calculation N randomised n for primary analysis Power to detect originally specified difference (%)a
Detectable difference with:
80% power 90% power
Original 472 400 90 14 percentage points
(30% vs. 44% = OR 1.84)
16 percentage points (30% vs. 46% = OR 2.00)
Revised 432 367 87 15 percentage points
(30% vs. 45% = OR 1.89)
17 percentage points (30% vs. 47% = OR 2.07) a Originally specified difference: 16 percentage points (30% response in UC vs. 46% in CBT = OR 2.00).
duration of the current depressive episode), which included adjustment for the baseline measurement of
the outcome (BDI-II score as a continuous variable). The ORs of‘response’ in the intervention group
compared with the usual-care group is presented along with a 95% confidence interval (CI) and the
p-value for the comparison.
Secondary analyses
Secondary analyses of the primary outcome included additional adjustment for any prognostic variables demonstrating marked imbalance at baseline (ascertained using descriptive statistics). The BDI-II was also considered as a continuous outcome, with an associated increase in power. Secondary analyses were conducted for other outcomes measured at 6 and 12 months.
Repeated measures analyses were used to incorporate the outcome data from both 6 and 12 months post randomisation (or, in the case of the PHQ-9, from 6, 9 and 12 months) to examine whether or not any treatment effects were sustained or emerged later. This was tested formally by the introduction of an interaction between treatment group and time. In the absence of any time effect, repeated measures analyses generate an average effect size over the duration of follow-up. In all analyses, ORs (or regression
coefficients for continuous outcomes), with 95% CIs and p-values, are reported.
Potential clustering by therapist
There is the potential for clustering by therapist within this trial, although clustering effects will operate in only one arm of the trial. Secondary analyses therefore used generalised linear and latent mixed models to
obtain a fully specified heteroscedastic model (described in Roberts and Roberts71) to examine the
influence of clustering by therapist on the results.
Subgroup analyses
Two subgroup analyses were specified a priori and were conducted by introducing an appropriate
interaction term to the regression model for the primary outcome. This permitted investigation of any
differential effects of treatment on outcome according to two predefined factors: (1) patient expectation
of outcome (defined as three levels: ‘CBT would definitely help me’; ‘CBT would probably help me’;
‘I don’t know if CBT would help me/would probably not help me’) and (2) degree of treatment resistance
[six levels based on duration of symptoms (< 1, 1–2, ≥ 2 years) and past treatment with antidepressant
medication (yes/no)].
Sensitivity analyses to examine the impact of missing data
The pattern of missing data was investigated by identifying those variables recorded at baseline that were
associated with‘missingness’ of the primary outcome (BDI-II score) at p < 0.20 at either the 6-month
follow-up and/or 12-month follow-up. Sensitivity analyses were conducted using the method of MICE72to
impute missing data (Stata ice procedure, version 1.9.5, dated 15 April 2011). The imputation model included all of those variables that were part of the substantive ITT model, together with the variables
associated with missingness (as identified above) and all available measures of depressive symptoms (BDI-II
and PHQ-9) and anxiety (GAD-7), irrespective of statistical significance. The baseline CIS-R score was also
included in the imputation model as another marker of severity. Variables included in the imputation model were declared as continuous, binary, categorical or ordinal variables, as appropriate. The match procedure was used to handle non-normally distributed variables that could not be successfully transformed. In total, 25 data sets were generated and 10 switching procedures were used.
Treatment efficacy
Complier-Average Causal Effect (CACE) estimates73for those who were viewed as‘on track’ to receive the
full course of CBT treatment at the time of the 6-month follow-up (defined as having completed nine or
more sessions) were estimated using instrumental variables regression methods. As the primary outcome ‘response’ was a binary variable, a probit transformation was used, and the primary ITT analysis repeated on this transformed scale for comparison with the CACE estimates.
METHODS
18
Complier-Average Causal Effect estimates were also determined for the longer-term outcome at
12-months. For the latter, those who had received≥ 12 sessions were regarded as having received CBT in
line with the treatment protocol.
The CACE methodology compares the outcomes for those who‘complied’ with the intervention with a
similar group of‘would-be compliers’ from those randomised to usual care, thus avoiding the biases
inherent in crude per-protocol analyses. The original definition of a ‘complier’ included those where the
therapy goals were achieved in fewer than 12 sessions, as agreed between the patient and therapist.
However, in practice, this included individuals who had an‘agreed end’ of fewer than eight sessions, and
thus a stricter definition was adopted, as outlined above.