Comparing the Effects of Neoadjuvant versus Adjuvant Chemotherapy in Stage II-III, Triple-negative Breast Cancer: Illustration of Trial Emulation Using the Clone-Censor-Weight Method

Background: The target trial framework can guide the robust design of comparative effectiveness studies using observational data. Emulation conducted with the clone-censor-weight method (CCW) addresses immortal time bias introduced by misalignment of study eligibility, start of follow-up, and treatment initiation to estimate per-protocol treatment effects.

Objectives: To apply the CCW method for evaluating effect of timing of chemotherapy in relation to surgery for stage II-III, triple negative breast cancer (TNBC) using the National Cancer Database (NCDB).

Methods: We identified incident female, stage II-III, adult (18-75 years) TNBC patients from 2013-2016 using the American College of Surgeons NCDB. We first cloned each patient and assigned each clone to 1 of the 2 treatment strategies. The neoadjuvant chemotherapy (NACT) strategy was defined as initiating chemotherapy within 9 weeks from diagnosis followed by surgery within 36 weeks from diagnosis. The adjuvant chemotherapy (ACT) strategy was defined as surgery within 9 weeks from diagnosis followed by initiating chemotherapy within 18 weeks from diagnosis. Clones were followed from diagnosis until death or end of data availability. Clones were analytically censored in their treatment arm when their observed treatment pattern deviated from the assigned strategy. For example, clones assigned to NACT were analytically censored at 9 weeks if they did not initiate chemotherapy within 9 weeks of diagnosis. To account for potentially informative censoring, we calculated inverse probability of censoring weights (IPCW) using 2 logistic regression models for each treatment strategy at 9 and 36 weeks for NACT and 9 and 18 weeks for ACT. We estimated 3-year mortality comparing NACT versus ACT using an IPCW-weighted pooled logistic model. The 95% confidence intervals were calculated using bootstrapping with 500 iterations.

Results: Our cohort included 36,241 women, resulting in a cloned dataset of 72,482 replicates. We observed 18,018 women who completed the NACT strategy, 9,179 women who completed the ACT strategy, and 9,044 women whose treatment did not follow either strategy in full. Median IPCW was 1.37 (Range: 1.00, 7.91; IQR: 1.00, 1.74) in the NACT group and 2.58 (Range: 1.00, 44.90; IQR: 1.00, 3.60) in the ACT group. Weights were truncated at the 95th percentile in the ACT arm. Age, histology, and treatment facility type had the largest imbalances before weighting, which improved after IPCW. Outcome results will be prepared for ICPE presentation.

Conclusions: Our study applied the CCW method to avoid immortal-time bias, a major concern in comparative effectiveness studies using observational data. Our results highlight key design and analytic decisions necessary for successful implementation of the CCW method.