1. Why Standard Regressions Fail with Time-to-Event Data
In oncology, surgical follow-ups, and chronic disease studies, clinical endpoints are rarely just binary ("Did the patient survive?"). The critical scientific question is "When did the event occur?"
If a patient in a 5-year cancer trial remains disease-free for 4 years before migrating abroad, standard logistic regression either misclassifies them as a failure or excludes them entirely. Survival analysis handles Censored Data by incorporating every patient's exact observation duration up to the moment they were last examined.
2. Understanding Censoring in Clinical Trials
- Right Censoring (Most Prevalent): Occurs when a participant has not experienced the study event by study close, is lost to follow-up, or dies from an unrelated event.
- Left Censoring: The event occurred prior to study entry, but the exact timestamp is unknown.
- Interval Censoring: The event is known to have occurred between scheduled screening appointments (e.g., between 6-month and 12-month follow-up CT scans).
3. Kaplan-Meier Estimation & Curve Anatomy
The Kaplan-Meier non-parametric estimator calculates conditional survival probabilities at each discrete point an event occurs:
1. Step-Down Visual Graph
The curve only drops at timestamps where an event occurs. Vertical tick marks represent censored observations.
2. Median Survival Time
The exact time on the X-axis where survival probability drops to 50% (0.50). If survival stays above 50% throughout follow-up, median survival is reported as "Not Reached (NR)".
3. Number-at-Risk Table
High-impact journals (e.g., JCO, Lancet Oncology) require a table aligned underneath the X-axis showing exact patients remaining at risk at regular intervals (e.g., 0, 12, 24, 36, 48, 60 months).
4. Cox Proportional Hazards Regression: Interpreting Hazard Ratios
While Kaplan-Meier compares single univariable factors using the Log-Rank Test, the Cox Proportional Hazards Model evaluates multiple clinical covariates (age, tumor stage, performance status, surgical margins) simultaneously:
- Hazard Ratio (HR) = 1.00: Identical event rate between arms across time.
- HR = 0.65 (95% CI: 0.48–0.88, p = 0.005): The treatment group experiences a 35% relative reduction in the hazard of death/recurrence at any given time point.
- HR = 1.80 (95% CI: 1.25–2.59): The predictor confers an 80% increase in risk.
Proportional Hazards Assumption:
Cox modeling assumes the ratio of hazards remains constant over time. Always test Schoenfeld residuals in SPSS/R. If hazard curves cross over time, use Restricted Mean Survival Time (RMST) or time-varying covariates.
Dr. Manjinder Singh Sidhu
MBBS, MD (Radiotherapy)
Senior Consultant in Radiation Oncology at DMCH Ludhiana with 18+ years experience, international fellowship at UCSD/Scripps USA, and 18 peer-reviewed publications.