The joint cross-sectional and time series characteristics of the data set could reasonably be expected to violate assumptions for the use of OLS regression, which necessitated the use of linear mixed-effects modeling [46]. With four repeated measures per subject and an a priori intraclass correlation coefficient (ICC) of 0.05, a sample size of 345 was calculated as sufficiently powered to detect an effect size of f² of 0.060, alpha level of 0.05, power of at least 0.80, and a maximum of 40 predictors. The data set, which was comprised of n=351 participants with measurements taken at four observation points (t=0, 6, 12 and 18 months) and limited to a set of 14 predictors drawn from our conceptual model, was sufficiently powered to support the use of LMM.

Test statistics (i.e., ICC, Akaike Information Criterion/Bayesian Information Criteria [AIC/BIC]) on the null model determined that a random intercept, random slope LMM model provided the best general fit across the SF-36 RP and RE outcome measures. Based on sample characteristics and test statistics we conducted our LMM modeling using maximum likelihood estimation and an unstructured covariance structure following procedures described in Rabe-Hesketh & Skrondal (2022) and Hoffman (2015) [47,48].

We determined the significance of a treatment group by time (and quadratic time) interaction term on our outcome variables in unadjusted LMM models. The unadjusted models were then used to construct trend line graphs of predicted values exhibiting the effect of treatment group on the RP and RE outcomes over the course of the 18-month study timeframe. We repeated this process for the aggregated SF-36 Physical Component Summary (PCS) and Mental Component Summary (MCS) outcome measures to provide a contextual background for our primary outcomes.

Our goal was to focus on the most predictive variables for our specific population that would facilitate clinical application, rather than all factors theoretically relevant in the Cancer Survivorship and Work Model. Therefore, we employed a stepwise backward deletion approach [49] to generate parsimonious final models, retaining predictors significant at p≤.05 while accounting for treatment group.

The baseline LMM construction was a random linear time model. We evaluated linear and quadratic time as both fixed and random effects [47]. Treatment group membership (i.e., CFAI, AI, controls) served as a time-invariant proxy for treatment characteristics, with the control group used as the reference category in model construction. Treatment group was retained across all models regardless of significance due to its centrality in comparing differences in role function among study participants. Pre-therapy time-invariant sociodemographic and health/treatment-related predictors [Online Resource Figure 1b] were added to the models during the first iteration of backward stepwise deletion. Accounting for significant time-invariant predictors, we added time-varying predictors related to symptom burden [Online Resource Figure 1b], again using stepwise backward deletion.

The treatment group unadjusted trend lines for RP, RE, PCS and MCS are presented in Figures 1 a–d. The trend line graphs depict the relationship between treatment protocol across time and each outcome variable. A significant positive treatment group by time interaction was found only in the RP (p<.001) and MCS (p=.002) models. The MCS trend lines displayed a partial downward sloping curvature. We include the unadjusted models for the summary scores of PCS and MCS solely as background context for RP and RE.

Turning to our primary outcomes of RP and RE, the final model results are presented in Table 2 (RP) and Table 3 (RE). A summary of the hypothesized and estimated coefficient direction and significance for the predictors for the final RP and RE models is presented in Online Resource 2. RP displayed significance with elements of both (a) sociodemographic and health/treatment-related time-invariant predictors in addition to the (b) symptom-related time-varying predictors. RE, in contrast, captured significant associations solely with respect to the symptom-related time-varying predictors. The treatment group by time interaction term in the final models was significant solely with respect to RP. RE showed no similar association.

The final RP model, which accounts for both fixed and random effects, showed a significant improvement over the OLS model without random effects (p<.001). The ICC in this final model was 0.45, indicating that 45% of the total variance in RP scores was attributable to differences between participants. This ICC was lower than the ICC of 0.61 observed in the null model, suggesting that the final set of predictors explained a portion of the between-participant variance. Most of the variance in the model was due to differences in participants’ mean baseline levels of RP functioning (random intercepts) rather than differences in how their functioning changed over time (random slopes).

The final model indicated several factors associated with RP limitations. Women with breast cancer receiving treatment, either CFAI or AI only, reported greater limitations compared to controls (both p<.001). The group by time interaction for both CFAI (p<.001) and AI only (p=.001) indicated that RP limitations decreased over the first 18 months of systemic therapy relative to the control group. Employed women reported fewer limitations (p=.004), while women with more comorbidities reported greater limitations (p=.003). Controlling for these baseline factors, we found that greater self-reported depressive symptoms (p=.006), pain interference with function (p<.001), cognitive problems (p<.001), and fatigue/inertia (p<.001) were associated with greater limitations in RP functioning. Conversely, greater self-reported tension/anxiety was associated with fewer limitations (p=.016).

The final RE model showed a significant improvement over the OLS model without random effects (p<.001). Most of the variance in the model was also due to differences in participant intercepts as opposed to variance due to their slopes. As with RP, clustering by participant as measured by the ICC (0.42) remained strong, although somewhat reduced, in relation to the ICC (0.54) of the null model.

There were no significant demographic and health/treatment-related characteristics in the final RE model. With the addition of symptom burden predictors in the second iteration, we found that greater self-reported depressive symptoms (p<.001), pain interference with functioning (p=.007), cognitive problems (p<.001), and tension anxiety (p<.001) were associated with greater limitations in RE functioning.

The time-invariant sociodemographic predictors used in our analysis were age, education, employment status, and married/partnered status. Only having employment was associated with better RP functioning. No significant sociodemographic predictors were found for RE functioning. Our finding is consistent with Mahar et al. (2008), who reported a significant positive association between BCS’ full-time working status and overall physical health using the SF-12 in a cross-sectional study [50]. Multiple studies have supported a positive association between work and HRQoL, of which role function is a component, in both general cancer survivor populations [51,52] and BCS specifically [50,53,54].

BCS’ employability hinges on understanding how AI treatment intersects with a woman’s life stage, a factor included in the Cancer Survivorship and Work Model. Notably, Mahar et al. (2008) reported a mean age of 48.8 [7.6] at diagnosis, while our study’s mean ages were 59.7 [5.17] in the CFAI group and 62.3 [SD 6.23] for the AI only group. Most of the women in our study were still within working age, albeit at an age more likely to encompass retirement planning. The psychosocial experiences of older versus younger BCS are not fully understood [55], including possible differences in perceived employment vulnerabilities [56]. Our findings suggest perceptions of diminished role function that may introduce added employment vulnerability. Understanding the unique circumstances of retirement-age BCS is key to preventing unplanned early departure from work [57] or promoting work reentry later in life [26]. Early in systemic therapy, when RP functioning is lowest (see Figure 1a), may be an opportune time for interventions to improve role function.

Health/treatment-related predictors examined in our study included the number of comorbidities and treatment group. Having more comorbidities was associated with lower RP functioning. Previous studies have found similar associations across different age ranges: Fu et al. (2015) reported that having more comorbidities was associated with lower SF-36 scores in four of eight domain scales at 12 months post-diagnosis in a relatively younger population with a mean age of 58.8 [11.8] [58], while Ganz et al. (2003) reported that having more comorbidities was associated with lower SF-36 Physical Functioning scale scores at 15-months post-diagnosis in older women age 65- ≥80 years [59]. Ganz, however, did not assess the SF-36 RP scale. Fu’s assessment included the entire set of eight SF-36 scales. Similar to Ganz, Fu found that greater number of comorbidities was associated with greater limitations in the Physical Functioning scale, which represents levels and kinds of limitations in physical activities ranging from easy to difficult [5], but did not report significance for the SF-36 RP scale that was the focus of our study. Overall, our studies suggest that comorbidities affect different aspects of HRQoL across a broad age spectrum of BCS during systemic therapy. In fact, a recent meta-review highlighted the need for further investigation of comorbidities’ impact on BCS’ psychosocial experience during systemic therapy [60]. Given our findings, perceived role function may be a key consideration in older BCS with greater comorbidities during the initial phase of systemic therapy.

Regarding treatment, we found that both CFAI and AI only groups were associated with worse RP functioning relative to controls, with CFAI having twice the impact on RP functioning compared to AI only. In contrast, a systematic review of thirteen randomized clinical trials (RCTs) conducted by Xiao et al. (2018) found that long-term use of ET (≥5years) was unlikely to adversely affect HRQoL in postmenopausal women [61]. Multiple differences may explain this discrepancy, including different ET agents, timeframes, and HRQoL instruments across studies. Furthermore, HRQoL outcomes were secondary and may have been underpowered. Only two RCTs in the Xiao review had similar treatment groups to our study, but these were extensions of earlier studies in women who had completed chemotherapy well before the extension [62,63] and one was in women who were older (≥70 years) [62] relative to the participants in our study. Future examination of survival relative to HRQoL outcomes such as role function is warranted to guide treatment decisions in working age postmenopausal BCS.

Our finding of a significant treatment group by time interaction indicates that the negative effect of treatment on RP diminished over time. While RP functioning for the CFAI group was lower than that of the AI only group at baseline, the trend lines converged by 18 months. Previous research in postmenopausal BCS found that HRQoL trend lines were significantly lower for chemotherapy recipients relative to an ET only group at 3 months [64]. While the treatment groups in that study are not directly comparable to ours, consistent with our study the treatment group HRQoL trends improved over time with no significant differences at 18 months [64]. Although we found no meaningful difference between CFAI and AI only groups at 18 months (Cohen’s d=0.04), we observed moderate effect sizes for differences between each treatment group and the control group (CFAI: Cohen’s d=0.54; AI only: Cohen’s d=0.43) that are potentially clinically meaningful [65–67].

While not the focus of our analysis on role functioning, it is worth noting that physical health component summary (PCS) scores remained below the threshold for impaired functioning [5] throughout the study (Figure 1c). The improvement in RP functioning over time, contrasted with consistently impaired PCS scores, suggests that BCS may compensate using coping strategies and problem-solving skills. We posit this consideration based upon the Cancer Survivorship and Work Model’s portrayal of a bidirectional influence between an array of factors in two of its domains; the Individual and Interpersonal domain in which coping and problem-solving are included and Effects of Cancer and Treatment domain in which role functioning is contained. More research is needed to understand whether BCS deploy coping mechanisms in order to maintain physical role function despite early adverse treatment effects and if such performance is sustainable especially if poor perceptions of overall physical health persist.