Potential mechanisms of the fatigue‐reducing effect of cognitive‐behavioral therapy in cancer survivors: Three randomized controlled trials

Abstract Objective Fatigue is a common symptom among cancer survivors that can be successfully treated with cognitive‐behavioral therapy (CBT). Insights into the working mechanisms of CBT are currently limited. The aim of this study was to investigate whether improvements in targeted cognitive‐behavioral variables and reduced depressive symptoms mediate the fatigue‐reducing effect of CBT. Methods We pooled data from three randomized controlled trials that tested the efficacy of CBT to reduce severe fatigue. In all three trials, fatigue severity (checklist individual strength) decreased significantly following CBT. Assessments were conducted pre‐treatment and 6 months later. Classical mediation analysis testing a pre‐specified model was conducted and its results compared to those of causal discovery, an explorative data‐driven approach testing all possible causal associations and retaining the most likely model. Results Data from 250 cancer survivors (n = 129 CBT, n = 121 waitlist) were analyzed. Classical mediation analysis suggests that increased self‐efficacy and decreased fatigue catastrophizing, focusing on symptoms, perceived problems with activity and depressive symptoms mediate the reduction of fatigue brought by CBT. Conversely, causal discovery and post‐hoc analyses indicate that fatigue acts as mediator, not outcome, of changes in cognitions, sleep disturbance and depressive symptoms. Conclusions Cognitions, sleep disturbance and depressive symptoms improve during CBT. When assessed pre‐ and post‐treatment, fatigue acts as a mediator, not outcome, of these improvements. It seems likely that the working mechanism of CBT is not a one‐way causal effect but a dynamic reciprocal process. Trials integrating intermittent assessments are needed to shed light on these mechanisms and inform optimization of CBT.

text). In the Gielissen-trial, a previous version of the CIS-activity has been administered which differs from the recent version in the wording of the three items; the scoring is identical. Higher scores indicate more perceived problems with activity (range [3][4][5][6][7][8][9][10][11][12][13][14][15][16][17][18][19][20][21]. Sleep disturbance was assessed with the subscale sleep and rest of the Sickness Impact Profile-8. 10 Seven items (e.g. 'I sleep or nap during the day') assess patients' functional impairment in daily life regarding sleep and rest. Patients are instructed to check those items that are applicable to them on a given day and are related to their health. Scores of the checked items are weighed. Higher scores indicate more limitations in sleep and rest (range 0-499).
Problems with social support were assessed with the discrepancy subscale of the Sonderen Social Support Inventory. 11 Eight items (e.g. 'What is your opinion about the extent to which people: Stand by you?') assess discrepancies between the amount of received and the amount of desired social support. Items are scored on a 4-point Likert scale ranging from (1) 'I miss it' to (4) 'It happens too often'. Higher scores indicate a higher discrepancy in support (range 8-32).
In the Abrahams-and Prinsen trial, depressive symptoms were assessed with the depression subscale of the Hospital Anxiety and Depression Scale. 12   Graded activity  Targets a fluctuating or low (physical) activity pattern  Patients with a fluctuating activity pattern learn to evenly distribute their activities during the day and will subsequently gradually increase their daily activity level (e.g. walking, cycling)  Patients with low activity pattern immediately start with gradual increase in their daily physical activity (e.g. walking, cycling) Realizing goals  Patients make an action plan to realize their formulated interventions goals  Patients learn to let go of the regular sleep-wake rhythm and even distribution of activities  Patients evaluate their progress The effect size Cohen's f 2 was calculated as ( 2 1+ 2 ). The direct effects are computed among post-treatment variables based on the causal model in Figure 2 (main text). f 2 ≥ 0.02 indicates a small effect, f 2 ≥ 0.15 indicates a medium effect, f 2 ≥ 0.35 indicates a large effect. 14 SE = Standard error.    The values in this table provide the degree of confidence in each mediation path, defined as the reliability computed by the causal discovery algorithm BCCD (bayesian constraint-based causal discovery) for this path, averaged over 1000 runs of this algorithm on half-sampled datasets. This definition leads to conservative values which do not sum up to one (100%) since we only considered two out of many alternative mediation models. Fatigue as an outcome refers to a mediation path of: Condition  Putative mediator  Fatigue. Fatigue as a mediator refers to a mediation path of: Condition  Fatigue  Putative mediator.  Polychoric correlation was used to compute the correlation between the dichotomous variables Condition and Sex. Polyserial correlation was used for pairs of either Condition or Sex with other variables. Pearson correlation was used for the remaining variable pairs. † indicates variables to which z-score transformation has been applied, consequently their mean values = 0 and SDs = 1. * indicates a significant correlation (p < 0.05). pre = preassessment, post = post-assessment. 9 S2b Means, Standard Deviations [SDs] and correlation matrix showing the pairwise correlations between the different variables without z-score transformation. Polychoric correlation was used to compute the correlation between the dichotomous variables Condition and Sex. Polyserial correlation was used for pairs of either Condition or Sex with other variables. Pearson correlation was used for the remaining variable pairs. * indicates a significant correlation (p < 0.05). pre = pre-assessment, post = post-assessment.

FIGURE S3
Classical mediation model: Sensitivity analysis without z-score transformation. Note: Even though the path from Fatigue catastrophizing to Fatigue is insignificant, the indirect effect from Condition to Fatigue through Fatigue catastrophizing is significant (ab = -2.06, CI[-4.68, -0.07]). * indicates a significant path.

FIGURE S4
Causal discovery sensitivity analysis without z-score transformation.
Values in which different questionnaire(-versions) have been used are treated as missing values. The tail (-) represents the origin of the causal effect and the arrowhead (➤) the direction of the causal effect. The circle (o) represents an association in which the origin and direction are unclear. The undirected lines () indicate the presence of selection bias (i.e. bias introduced by the sample selection). All links represent a causal association of which the edge has a post-bootstrap reliability coefficient of ≥ 0.5, with a thicker line corresponding to a more likely causal association between variables. The values represent the strength of the causal effects (see also Table S2b).