Accelerated failure time as an alternative to Cox proportional hazards for parathyroid carcinoma in the Surveillance, Epidemiology, and End Results database.

Abstract

e18139 Background: Cox proportional hazards regression is a common model for multivariate survival analysis. This model relies on proportional hazards (PH), meaning hazard ratios for predictor variables must remain constant over time. Failure to report PH violation testing is prevalent in published studies; consequences of violation include decreased power and unreliable modeling. Large databases such as the Surveillance, Epidemiology, and End Results (SEER) cancer database allow clinicians to draw conclusions regarding treatment for rare cancers such as parathyroid carcinoma (PC), as there are no large-scale studies. Cox regression is commonly used to analyze this database. Multiple studies advocate for alternative models such as Accelerated Failure Time (AFT) when PH is violated; however, no studies have assessed PH violations or used the AFT model specifically for PC. The goal of this study is to assess PH violations in the SEER database for PC patients and demonstrate utility of AFT as an alternative to Cox regression. Methods: We conducted a systematic review of three databases to identify papers reporting multivariate survival analyses of PC patients in the SEER database. We extracted individual variables used for multivariate analysis in the included papers and used these variables to independently analyze patients in the SEER database diagnosed with PC from 2002 to 2022. PH violation was assessed via visual inspection of univariate Kaplan-Meier curves. Multivariate analysis was conducted using Cox regression and AFT models, and the results of these two models were compared. Results: Our systematic review returned 783 results, and 14 studies met criteria for inclusion. Most studies (92%, n=13) reported Cox regression results, but none of the 13 papers reported any testing for PH violation. We extracted 14 unique predictor variables from these papers. We then independently identified 666 PC patients in the SEER database; in this subset, a majority of the 14 variables violated PH in univariate analysis of overall (79%) and disease-specific (86%) survival. There were notable differences between the Cox regression and AFT models: simple (p = 0.046), complete (p = 0.02), and radical (p = 0.003) surgical intervention; negative lymph node metastases (p < 0.001); and above-average income (p = 0.04) were significant predictors of longer disease-specific survival in the AFT model but not in Cox regression. Other significant predictors of shorter disease-specific survival in the AFT model included Black race (p = 0.001), age > 72 years (p = 0.004), tumor size > 35mm (p < 0.001), regional disease (p = 0.047) and distant disease (p < 0.001). Conclusions: AFT is a more appropriate multivariate statistical model than Cox regression for PC patients in the SEER database. The AFT model identifies several clinically relevant survival predictors in this patient subset.