OBJECTIVES: Directed Acyclic Graphs (DAGs) are structured diagrams describing the assumed underlying causal relationships between treatments, outcomes, and covariates. As evidence generation using real-world data (RWD) continues to expand, DAGs can help identify biasing paths, inform selection of covariates, and generate a more valid causal effect estimate. This study systematically built DAGs to elucidate causal pathways between first-line (1L) treatment (immune-oncology (IO) versus non-IO therapy) and survival among patients with metastatic colorectal cancer (mCRC).

METHODS: The study design comprised three components: 1) two targeted literature searches to identify relevant variables related to treatment (IO [nivolumab, ipilimumab, or pembrolizumab] versus non-IO [standard of care chemotherapy]) and outcomes (progression-free/overall survival); 2) plausibility assessments of each proposed relationship to be captured in the DAG, and 3) a case study utilizing Carelon Research’s Healthcare Integrated Research Database (HIRD). Using US integrated claims and clinical data from 2014-2023, we differentiated measured from unmeasured variables and estimated bivariate relationships between each covariate, treatment, and overall survival via logistic regression to refine the proposed relationships.

RESULTS: The two targeted literature searches identified 94 RCTs and 22 RWD studies, from which 28 variables were extracted. These potential confounders (e.g., tumor characteristics, performance status, health care access) or colliders (e.g., data collection methods) relative to the treatment-outcome relationship were built into the DAG. We identified 9,046 patients with mCRC from the HIRD, of whom 213 patients initiated IO therapy (mean age 60 years, 47% female, 80% Non-Hispanic White). Age, performance status, prior chemotherapy, and BRAF mutation status demonstrated large effect sizes and statistically significant associations with both treatment (IO versus non-IO) and overall survival.

CONCLUSIONS: Creating DAGs in a systematic and efficient manner, informed by existing literature and plausibility assessments, provides transparency when estimating causal effects from RWD and can reduce bias in the chosen statistical model.