Speaker
Description
Survival is commonly used as a primary endpoint in preclinical cancer studies to assess treatment effects across different dose levels. Using a preclinical immuno‑oncology study as a case study, we examine the limits of survival‑based dose‑response inference. Cox and AFT models indicate a statistically significant dose‑response effect. However, key dose‑response parameters show substantial uncertainty. Using profile‑likelihood and Fisher‑information methods, we show that this uncertainty stems from having too few doses in the transition region of the dose‑response curve, indicating a limitation of the experimental design rather than the model itself.
Motivated by this, we use a pseudo‑Bayesian design approach to examine how dose placement could be improved in future studies. To complement the survival‑based analysis, we analyze longitudinal tumor trajectories. RECIST‑based stratification reveals a bifurcation‑like response pattern, with tumors showing either sustained regression or rapid escape growth, which is not apparent from survival analysis alone. Aggregate measures such as tumor growth area under the curve confirm modest overall treatment effects but do not explain this separation.
Overall, while survival analysis effectively detects treatment effects, its ability to characterize dose‑response relationships is limited by dose placement and censoring. Moreover, key dynamic response patterns remain hidden without additional trajectory‑level analyses.
