The invisible force that shapes every survival curve
You enroll 200 patients in a study of graft patency after lower extremity bypass. At 3 years, 40 patients have had graft occlusion, 30 transferred care or were lost to follow-up, and the study ended before the remaining 130 reached 3 years of follow-up.
What is the 3-year patency rate? It is not 80% (160/200). You do not know what happened to those 160 patients. Some never reached 3 years. Some disappeared.
Censoring is what happens when you do not observe the event for a patient -- not because it did not happen, but because you stopped watching. How you handle censored observations determines whether your survival analysis is valid or garbage.
Not all censoring is created equal. The type determines how much trouble you are in.
The event has not occurred by the time observation ends. You know the patient survived at least until the last follow-up, but not what happened after.
Patient had a patent graft at their last clinic visit (month 18), then moved out of state. You know the graft lasted at least 18 months.
The event occurred before observation began. You know it happened but not exactly when.
A patient presents with an occluded bypass graft. You know it failed at some point after surgery, but you do not know whether it was at 1 month or 11 months.
The event occurred between two observation points. You can bracket the timing but cannot pinpoint it.
Duplex ultrasound at 6 months shows a patent stent. Duplex at 12 months shows occlusion. The event happened somewhere between months 6 and 12.
The most benign form. The study ended on a fixed date, and some patients had not yet had the event. This is expected and usually non-informative -- the study end date has nothing to do with patient outcomes.
Standard Kaplan-Meier analysis handles right censoring. Left and interval censoring require specialized methods. If your data has substantial left or interval censoring and you use standard KM, your results are biased.
This distinction is everything. It determines whether your survival estimates are trustworthy.
Limb preservation studies are especially vulnerable to informative censoring. Patients with critical limb-threatening ischemia (CLTI) have high mortality. If a patient dies before amputation, they are censored for the limb salvage endpoint -- but death and amputation are not independent events. Sicker patients are more likely to both die and lose their limbs.
This is why competing risks analysis exists (Lesson 5).
Mistakes that even experienced researchers make.
"We excluded patients lost to follow-up from the analysis."
This introduces survivorship bias. You are keeping only patients who stayed in the system long enough to observe their outcomes. If sicker patients were more likely to be lost, your results will be overly optimistic. Censoring, not exclusion, is the correct approach.
"Follow-up was 100% complete."
Impressive but rare. And patients who died of other causes before the event are still censored. Even with perfect follow-up, censoring from competing events remains. The claim usually means no patients were lost, not that there was no censoring.
"Mean follow-up was 3.2 years."
Mean follow-up is nearly useless. It can be dragged up by a few patients followed for 10 years while most were followed for 1 year. Report median follow-up with the range or IQR instead.
"We treated censored patients as event-free."
Worst-case bias. A patient censored at 6 months is not event-free at 5 years. They had an unknown outcome. Treating them as event-free inflates survival estimates. This is one of the most common errors in retrospective surgical series.
When reviewing a paper, always ask: What was the censoring rate? Was it balanced between groups? What were the reasons for censoring? If these questions are not addressed, the results may be unreliable.
Identify the censoring type and its implications.
Module 4 - Lesson 2 complete
Next up: The Log-Rank Test -- how to formally compare two survival curves instead of just eyeballing them.