Association of sociodemographic, geographic, and health factors with survival across randomized cancer trials.
TL;DR
Sociodemographic, geographic, and health factors explained more than half of the prognostic information associated with overall survival across randomized oncology trials, exceeding that attributable to traditional clinical factors.
Abstract
260 Background: Clinical prognostic factors are routinely reported in oncology trials, whereas sociodemographic (SD) and geographic variables, when reported, are typically intended to reflect trial access and representation. Quantifying their association with overall survival (OS) may clarify the scientific value of their systematic collection and reporting. Methods: We analyzed randomized oncology trials using Cox proportional hazards models to quantify OS variation attributable to prespecified covariate domains. Attributable variation (AV) was estimated using likelihood-ratio methods. For each trial, domain-level AV was calculated from the increase in -2 log likelihood after removal of all covariates within a domain from the full multivariable model. Treatment assignment was retained in both the full and reduced models to assess the prognostic contribution of each domain independent of treatment effects. Domains included clinical (study-specific clinical prognostic variables used for stratification), SD (age, sex, race, ethnicity, area-level deprivation, and insurance status), geographic (US region and rural/urban status), and health status (performance status and obesity) factors. A similar likelihood-based approach was used to estimate AV for individual factors. Domain- and factor-level AV estimates were pooled across trials using weights proportional to each trial’s total prognostic information. Uncertainty in pooled AV estimates was assessed via trial-level bootstrap resampling. Results: In 85 clinical trials enrolling 44,365 patients (1992-2024), clinical risk factors were most commonly related to stage and disease characteristics (47.7%) and prior treatment (18.8%). Clinical factors accounted for the largest portion of treatment-adjusted AV in OS (40%; 90% bootstrap IQR (BIQR), 31%-46%), followed by SD (31%; BIQR, 27%-36%), health status (15%; BIQR, 12%-18%), and geographic factors (8%; BIQR, 6%-9%). Within the SD domain, age contributed most of the AV (52%), followed by sex (12%), area-level deprivation (12%), race (11%), insurance (7%), and Hispanic ethnicity (7%). Performance status accounted for most health-domain AV (78%); geographic AV was primarily driven by region (72%) followed by rural/urban status 28%. Clinical domain AV was distributed across multiple disease-specific prognostic factors, reflecting trial heterogeneity. Conclusions: Sociodemographic, geographic, and health factors explained more than half of the prognostic information associated with overall survival across randomized oncology trials, exceeding that attributable to traditional clinical factors. These findings provide an important scientific rationale for the routine collection and reporting of these variables as core components of cancer clinical trials, given their importance for interpreting outcomes and assessing generalizability.