Statistics and power analysis
Primary Aim: The primary aim of this study is to investigate the differential effects of different modes of acupressure on fatigue in breast cancer survivors. For this, we are focusing on three fatigue outcomes, namely, BFI, which will be administered weekly, a measure of average daily physical activity and a self-reported fatigue score, both based on Actiwatch-S data. A BFI is collected weekly and provides a measure of average degree of fatigue, a score that ranges between 0 and 10, with higher numbers indicating more fatigue. We will be analyzing the BFI outcome over phases. The first phase BFI is obtained by averaging week-1 to week-3 measures (mid-treatment). The second phase comprising of week-4 to week-6 measures indicates the status in the second half of the treatment phase (end-treatment). The final phase comprises of week-7 to week-10, the washout phase (washout). In order to investigate the change in BFI between the three study arms, we are using a mixed-effects regression analysis with average BFI in the three phases as the outcome, phase as the within-subjects factor, and group (RA, SA, standard of care) as the primary between-subjects factor. The baseline BFI level will be used as a covariate. We shall further adjust the model for key demographic variables such as age and ethnicity, as well as the baseline HADS score. A subject random-effect will be used to account for the clustering effect within the three measurements on the same subject. Of primary interest is the group*phase interaction, the significance of which will indicate differences in extent of decrease between the arms. Post-hoc analyses of phase difference within each arm will be carried out multiple comparison adjustment. Model diagnostics will be carried out to confirm the distributional assumptions and appropriate corrective actions (e.g. transformation) will be performed as needed.
Two types of fatigue-related measures will be collected from the Actiwatch-S. The first is a daytime activity count per minute, which is calculated by dividing the total daily activity counts by the total daytime minutes. For each phase as defined above, we shall use the corresponding 21-day (or 28-day) average as appropriate. The second type of measure is obtained from the four self-reported readings of the VAS-F. The daily averages are obtained from the non-missing entries each day, and are then combined to provide the phase average the same way as the daytime activity count. The analytical framework is identical to that for BFI, albeit with the changed outcome.
A secondary analysis with BFI will also be carried out which investigates the difference in proportion of severely fatigued subjects (BFI ≥ 6) across the three study arms. The analytical framework will be that of a mixed-effects logistic regression model with severe BFI (yes/no) being the dichotomous outcome. Apart from group, phase, and group*phase interaction as primary covariates, the model will also be adjusted for baseline status of severity. Remaining covariates in the model are identical to those in the continuous BFI model. The clustering effect due to subject will be accounted for using a generalized estimating equations approach.
In order to assess the extent of association between the improvement in FACT-B and improvement in fatigue measure (specific aim #1 letter D), we shall create the change-scores (week 6 – baseline, week 10 – baseline) for FACT-B (both total and component scores), and BFI. Then we shall fit a linear mixed model similar to that for the original BFI analysis, with BFI change-score as outcome, and the FACT-B change-score as a time-varying covariate. There will not be any phase or week variable in the model. We shall restrict this analysis to the SA and RA arms only so that the group variable will consist of two levels. A group*FACT-B interaction in the model will investigate any differential association across the arms between the change-scores. The additional covariates and clustering adjustment would be similar to the BFI analysis.
Secondary Aims:
1.
Assessing difference across study arms with respect to sleep parameters
One of the most widely used sleep measures is the PSQI, which shall be administered at baseline, week 6 and week 10. We shall follow the scoring convention followed for PSQI which constructs scores for seven components, namely sleep quality; sleep latency; sleep duration; habitual sleep efficiency; sleep disturbances; use of sleep-promoting medication; and daytime dysfunction, all derived from the original PSQI. The component scores are summed to obtain a total score, the primary outcome for the proposed study. The modeling framework and subsequent analysis will be similar to that for BFI with the phase variable replaced by a continuous week variable. A logistic regression analysis will be carried out with a dichotomous outcome indicating whether PSQI > 7 or not since the threshold of 7 is considered an indicator for sleep quality in BC survivors. A number of secondary sleep outcomes will be obtained using a conjunction of the Actiwatch-S-S data and sleep diary. These will include measures of total sleep time (minutes spent asleep each night) and sleep efficiency (total sleep time/time in bed* 100). These can all be analyzed using the models and methods described above within the mixed linear or logistic regression modeling framework as appropriate.
2.
Assessing differences in time to onset and time to relapse (after treatment ends) across the acupressure arms
Time to onset of the acupressure effect will be investigated by studying the treatment subgroups, namely RA and SA. For fatigue, we shall use the daily average values of VAS-F. Time to onset will be calculated as the first time (in days) the VAS falls down by 3 points in comparison to the baseline. We shall restrict this analyses to subjects whose baseline VAS-F is at least 3. A proportional hazards regression model will be employed to analyze the time to onset data. Treatment arm (RA vs. SA) will be used as the primary covariate in the analysis. Independent variables used in the earlier regression analyses for the primary and secondary aim 1 will be used as additional covariates.
For sleep, we shall use the VAS-SQ measure to carry out our time to onset analysis. As in the case of fatigue, a 3-point drop from baseline in VAS-SQ for the first time will be considered as onset. Since the VAS-SQ measures are weekly, we shall employ a discrete survival analysis technique to analyze this data. A discrete survival model is a logistic regression model of the discrete-time hazards, and is easily implemented in most standard statistical softwares. As in the case for VAS-F, this analysis will be confined to subjects with a baseline VAS-SQ score of 3 or more.
We would analyze similarly the time to relapse since the end of treatment, which is defined as the first time a 3-point increase (VAS-F or VAS-SQ) from the six-week value, occurs. This part of the analysis will be restricted to subjects for whom an onset has occurred.
All statistical analyses will be carried out in SAS 9.3 and PASW version 19.
Power analysis
Our power analysis is based on our primary aim of comparing fatigue as measured by BFI across the three study arms.
We compute the power via simulation using a mixed effects model with a between-subject factor group (3 levels), a within-subject factor phase (2 levels), group*phase interaction and a random subject effect. The mean BFI values at mid-treatment were assumed to be 4, 3, 3, respectively for standard care, SA and RA arms whereas the means were taken to be 4, 3, 2 at the end-treatment point in the same arms. The between-subject variance is assumed to be 4 at all time-points whereas the variance of the random subject component is taken to be 4 also (yielding an intra-class correlation of 0.5). These assumed values are estimates based on our pilot data. For this configuration, the power for detecting the difference between groups is more than 0.95 and the power for detecting a significant phase*group interaction is 0.82 with a sample size of 100 per treatment arm and a 5% level of significance. In the second model discussed in the analysis section with phase comprising of end-treatment and washout, if we assume the washout point BFI values to be 5, 4, 3, in the standard care, SA, and RA arms, respectively, the power for detecting either the phase effect or group effect are both around 99%.
We also have powered the study to allow us to observe differences in sleep parameters. The extent of overlap between fatigue and sleep disturbances in BC survivors is currently unknown. Research indicates that 51% of BC survivors experience sleep disruptions and 19% meet diagnostic criteria for insomnia[
35]. While fatigued women are most likely enriched with individuals experiencing sleep disturbances we will conservatively assume that roughly 51% women will have some significant sleep disturbances and 19% will have insomnia per treatment arm. As such, with 100 women per treatment group we will be overpowered for detecting a difference in fatigue and sufficiently powered to detect changes in key sleep measures. For example, using the PSQI as a basis for powering differences in sleep quality from previous studies the mean PSQI in BC patients ranges from 6.84 ± 0.376 to 7.16 ± 0.325.(60,74) A 1 point decrease in the PSQI from ~7 to ~6 is considered clinically significant in this population[
33]. The mean PSQI values at mid-treatment are assumed to be 7, 6, 6, respectively in standard care, SA and RA arms whereas the means are taken to be 7, 6, 5 at the end-treatment point in the same arms. Assuming an intra-subject correlation of 0.5, we have powers above 99% to detect any of the group, phase, and group*phase interaction effects with a sample size of 100 per treatment arm and a 5% level of significance.
All power calculations are carried out on the basis of 100 simulations in the software PASS 2008 (NCSS, Kaysville, Utah, USA).
Data safety monitoring plan (DSMP)
The PIs review study progress weekly with study staff, and problems with or pertaining to study subjects are be communicated immediately. The entire research team meets monthly to review progress and any problems encountered. This team includes one physician, a PhD trained nurse and trained acupuncturist who are highly experienced with cancer control trials. The PIs are notified when an AE occurs and determine the attribution and relatedness of each adverse event. All AEs must be given to the PI within 48 hours if involving a death or life threatening event, or within one week, if serious (not-life threatening/death) or non-serious. Lastly, we work with the UM Prevention Research Base DSMB which meets monthly by means of regularly scheduled meetings.
Composition of the UM Prevention Research Base DSMB: the PI is present in an open session portion of the meeting and absent in a closed session. All DSMB official subjects in the review of confidential data and discussions regarding continuance or stoppage of a study have no conflict of interest and no financial stake in the research outcome. The current UM Prevention research base Data and Safety Monitoring Committee is Chaired by the Dr. Mack Ruffin and comprised of Faculty members from the departments of gastroenterology, Family Medicine and Hematology/Oncology. At least 3 faculty members, not including the study PI, must be present to have quorum. If the DSMB cannot meet face-to-face, a conference call is acceptable. Prior to each meeting, the UM Prevention Research Base clinical research associate distributes a standard summary report detailing accrual, new publications or presentations relevant to the ongoing project, quality control audit information, any ethical concerns, patient-subject complaints and adverse events or serious adverse events.
Our study was initially reviewed by The UM Prevention Research Base DSMB beginning in the second year from initial funding and monthly after that. The DSMB also reports its findings of any adverse events or decisions regarding modification of the protocol to the University of Michigan IRB committee.