br therapy or the only adjuvant therapy Furthermore given th
therapy or the only adjuvant therapy? Furthermore, given that the addi-tion of RT did not yield an OS benefit for the group as a whole, the pa-tients for whom this treatment strategy may yield the greatest benefit would most likely have significant risk for LRR while a relatively low competing risk of distant failure. This may reflect a patient population with a combination of high-risk features for LRR, such as deep myometrial invasion, serosal involvement, cervical invasion, or large tumor size, in addition to the absence of high-risk features for distant failure, such as high grade or non-endometrioid histologies. To test this hypothesis, the National Cancer Database (NCDB) was queried for women with locally advanced Type 1 EC who underwent hysterectomy and received adjuvant CT and/or RT. Survival analyses were performed to estimate the association of sequence of CT and RT with OS in propen-sity score-matched cohorts.
2. Methods and materials
2.1. Cohort selection
The NCDB (2004–2015) was queried for women with FIGO Stage III-IVA (pT3–4 or pN1–2) Type 1 (grade 1–2, endometrioid histology) EC following hysterectomy (Fig. S1). The NCDB is a nationwide, facility-based comprehensive clinical surveillance resource oncology dataset established by the Commission on Cancer of the American College of Surgeons and the American Cancer Society in 1989 that captures 70% of all newly diagnosed malignancies . The American College of Sur-geons has executed a Business Associate Agreement that includes a data use agreement with each of its Commission on Cancer accredited hospitals. Local institutional review board approval and informed con-sent were not required for analysis of de-identified data.
Women who received single-agent CT only, inadequate RT (b45 Gray (Gy), b25 fractions, or unknown) were excluded, as were women whose primary listed RT volume was other than pelvis or uterus. It was not possible to determine if pelvic and/or para-aortic CI-898 node volumes were included in the RT treatment plan. The se-quence of RT and CT was determined by documented treatment start dates (RT before CT: RT start date ≥30 days (d) before CT start date; RT after CT: RT start date ≥ 62 d after CT start date). Women with un-known sequence of RT and CT were excluded. To clearly define se-quence of RT and CT, women with a documented RT start date of 15–61 d following the start of CT were excluded, as were women for whom the start date for RT was N200 d either before or after the start date of CT (Fig. S2). To account for immortal-time bias, follow-up time was calculated from the start date of the patient's last initiated treat-ment. Covariates were included in the creation of the matched cohorts and to adjust for potential confounding during regression analyses (Supplemental Methods).
2.2. Statistical analyses
Baseline characteristics between patient groups were compared using the Fisher's exact test for categorical data, the Mann-Whitney U test for non-normally distributed numeric or ordinal data, or the t-test or ANOVA for normally distributed data. Five-year restricted mean sur-vival times (RMST) for OS and 5 y OS proportions were estimated using the Kaplan-Meier method and compared with the log-rank test. Sur-vival curves were plotted as unadjusted Kaplan-Meier estimates.
Multivariable parametric accelerated failure time (AFT) models using the generalized gamma distribution were used, as described pre-viously, to evaluate the association of adjuvant therapy modality and se-quence with OS . This model was chosen in place of the Cox proportional hazards (PH) model due to the presence of significant non-PH among women oils received CT only when the Cox model was used for the multivariable analyses . When the PH assumption is found to be violated using the Cox model, the AFT model for multivar-iable analyses provides better goodness-of-fit to the observed data and
therefore more robust statistical inference [6–9]. The AFT model esti-mates the time ratio (TR), which describes the multiplicative factor by which the time-to-event is related between two groups. A TR N 1 de-scribes longer survival. Covariates utilized in analyses were selected a priori based on clinical knowledge and availability and are described in detail in Supplemental Methods.
To reduce potential confounding, nearest-neighbor propensity score-matching was performed without replacement using a caliper size of 0.05 for each paired subset to generate well-balanced matched cohorts, as described in detail in the Supplemental Methods. Propensity score-matched and inverse probability-weighted cohort analyses were performed to reduce treatment-assignment biases related to measured covariates . Analyses were performed both for all patients as well as for the matched patient cohorts within each paired subset. All statistical tests were two-tailed with an alpha of 0.05 used as the cut-off for statis-tical significance. Sensitivity analyses were performed on an expanded cohort of women to also include Grade 3 endometrioid histology. To test the robustness of the observed results, sensitivity of TR estimates to a possible unmeasured confounder was explored [11,12]. Statistical analyses were performed in Rstudio v1.1.453 using MatchIt, survival, and flexsurv, packages [7,13–15].