Application of BRACE Method to Address Treatment Selection Bias in Observational Data

Abstract

Background: Cancer treatments can paradoxically appear to reduce the risk of non-cancer mortality in observational studies, due to residual confounding from treatment selection bias. Here we apply a novel method, Bias Reduction through Analysis of Competing Events (BRACE), to reduce bias in the presence of residual confounding. Methods: We studied 36630 prostate cancer patients, 4069 lung cancer patients, and 7117 head/neck cancer patients, using the Veterans Affairs Informatics and Computing Infrastructure database. We estimated effects of intensive treatment (prostate: prostatectomy vs. radiotherapy; lung: lobectomy vs. sublobar resection or radiotherapy; head/neck: radiotherapy with concurrent cisplatin and/or multiagent induction vs. radiotherapy with or without alternative systemic therapy) on cancer-specific mortality, non-cancer mortality, and overall survival (OS), using both multivariable Cox (MVA) and propensity score (inverse probability treatment weighting (IPTW)) models. Next, we applied the BRACE method to adjust for residual confounding, based on the observed treatment effect on competing event and relative event hazards. Results: For each cohort, intensive treatment was associated with significantly reduced hazards for cancer-specific mortality, non-cancer mortality, and OS. Compared to the results for MVA and IPTW models, hazard ratios (95% confidence intervals) for the effect of intensive treatment on OS were attenuated in each cohort after applying BRACE: (prostate- MVA: 0.75 (0.71, 0.80), IPTW: 0.73 (0.66, 0.75), BRACE: 0.98 (0.95, 1.00); lung- 0.79 (0.68, 0.91), 0.79 (0.66, 0.89), BRACE: 0.81 (0.65, 0.94); head/neck- 0.71 (0.66, 0.76), 0.70 (0.66, 0.76), BRACE: 0.81 (0.76, 0.86)). BRACE estimates were similar to findings from meta-analyses and randomized trials. Conclusions: We found evidence of residual confounding in several observational cohorts after applying standard methods, which were mitigated after applying BRACE. Application of this method could provide more reliable estimates and inferences when residual confounding is identified and represents a novel approach to improving the validity of outcomes research.