Primary analyses will be conducted on an intention-to-treat basis. Thus, results for all randomised women will be analysed in the group to which they were assigned, regardless of protocol violations. The only exception to this will occur if participants withhold or withdraw consent to use their data in the analysis. Statistical analyses will be conducted in R (data management, graphs, preliminary analyses) and MPlus using MplusAutomation [
75]. As a preliminary step, data will be assessed for outliers and skew and, where appropriate, transformed or winsorised. Prior to primary analyses, we will produce a thorough descriptive profile of the sample, characterising both the dropout and missing data. There are no interim analyses planned. Statisticians will not be blinded.
Baseline comparisons
Participant characteristics at baseline will be presented by treatment group, and statistical tests will be conducted to verify that randomisation was successful. Discrete variables will be summarised by frequencies and percentages, and baseline group differences tested using chi-square tests. Continuous variables will be summarised using mean (standard deviation) or median (interquartile range) and baseline group differences, tested using independent t tests or Wilcoxon signed-rank tests for variables that evidence non-normal distributions. If there are significant differences in any participant characteristics at baseline, the variable(s) will be included as covariates in the final analyses.
Aim 1
Aim 1 is to test the efficacy of CBT+ compared to TAU+ for improving symptoms of sleep disturbance. Aim 1 will be tested using LGMs. LGMs will be estimated with an intercept and two linear slopes, representing a piecewise model. Slope 1 will have loadings constrained to 0, 0.5, 1.0 and 1.0 for weeks 0, 3, 6 and 12, respectively, capturing the linear change from baseline (week 0) to post-intervention (week 6). Slope 2 will have loadings constrained to 0, 0, 0 and 1.0 for weeks 0, 3, 6 and 12, respectively, allowing a different slope from post-intervention (week 6) to follow-up (week 12) compared to from baseline (week 0) to post-intervention (week 6). The means and variances of the intercept and slope factors will be freely estimated (corresponding to random effects in linear mixed models) and the intercept and slope covariances will be estimated. The residual variance will be constrained to equality across time and residuals assumed uncorrelated, corresponding to an independent, homogenous residual structure. Intercepts of indicators will be constrained to 0 to allow estimation of the latent random intercept mean.
There are two stratification factors: screening ISI (≤7, ≥8) and cancer stage (≤2, ≥3). These factors will be crossed, creating four groups: early-stage, low ISI; advanced-stage, low ISI; early-stage, high ISI; and advanced-stage, high ISI. Dummy codes will be created for each strata, with early-stage, low ISI treated as the reference group. These dummy codes will be included as covariates to adjust for their effect on the random intercept following recommendations that stratification factors be adjusted for in analyses of randomised controlled trials [
76,
77].
Treatment effects will be evaluated by creating a dummy code (0 = TAU+, 1 = CBT+). This dummy code will be entered as a predictor of the intercept, slope 1 and slope 2. However, treatment factors will be constrained to 0 for the intercept, to implement so-called constrained longitudinal data analysis, which studies show provides a more accurate estimate of treatment effects from randomised controlled trials with repeated measures [
78,
79].
The primary trial results will be the effect of treatment on slope 1 (i.e. from week 0 to week 6). This interaction directly tests whether the change in primary outcomes over time is different in the control and intervention arm. Effect sizes of the group difference at each time point also will be calculated.
A similar process will be followed for sleep efficiency; however, a continuous time, linear model will be estimated using a mixed effects model in Mplus because up to 42 days of sleep efficiency are collected. A piecewise model is not needed as sleep diary data are not collected at follow-up due to burden. Sleep efficiency may not follow a normal distribution. If the assumption of normality is violated, significance tests and confidence intervals will be based on non-parametric bootstrapping.
Aim 3
Aim 3 is to explore potential mechanisms of CBT+ and moderators of the intervention efficacy. Mechanisms (i.e. pre-sleep arousal, beliefs and attitudes about sleep, vulnerability to insomnia under stress and intrusive thoughts before bed) will be tested using mediation conducted in path analyses. Specifically, we will examine the effect of condition on change in mechanisms (for example, pre-sleep arousal) from baseline to treatment mid-point, and test whether the change in mechanisms from baseline to treatment mid-point accounts for the condition effects on change in outcomes (for example, ISI) from baseline to post-treatment and follow-up. Statistical mediation will be determined by evaluating the indirect effects, calculated as the product of the paths from condition to change in mechanisms and from change in mechanisms to change in outcomes. Bootstrapping will be used to estimate the confidence interval for indirect effects and their statistical significance. Given the modest sample size, analyses will be conducted for each mechanism and outcome individually.
Treatment moderators (i.e. pain, chronotype, perceived credibility and expectations of the intervention, and adherence to the intervention protocol) will be evaluated by modifying the primary linear mixed effects analyses from aims 1 and 2 to include a condition × moderator interaction, along with all lower order effects predicting slope 1 and slope 2 of the piecewise model. Individual moderators will be tested in separate models.